As of my knowledge cutoff, three of the top state-of-the-art large language models are Google's Gemini 2.5 Pro, OpenAI's o3 (and GPT-4.5), and Anthropic's Claude 3.7 Sonnet.
I'm using this model via API. I am most excited about its ability to write the english language (or any language). My least favorite trend on twitter has been people posting 2D illustrations/3D renders as a proxy for model capability. Well-written prose is far less likely to go viral but far outstrips the importance rendering a fantasy world in one shot.
I've been trying it out via the API and the model has been a delight so far. Clear in the head, not RL-fried, distinct writing style, and big model smell. That said, the em-dash is back w/ this model.
That said, I would be so frustrated with Sundar Pichai if I was on the Google team. The botched rollout of this model seems to be entirely a matter of capacity, and Pichai has been more than happy to give it to anybody but his own team.
I've been using it for a while internally at Google. I prefer Argon over Opus 5.5 because of the clear thinking and directness. It phrases things in an organized, straightforward way.
I have used it as well. It is very dense, but it cites with evidence and has a direct and straightforward approach and manner. It doesn't really have a "tone" or mannerisms the way Opus 5 did. It follows instructions and sticks with pretty clean language by default. Argon is solid and handles long horizon well.
Announced the model two weeks ago to a ton of hype, model was unavailable for anyone to try. On twitter, Google folks went on a great run taking shots at Muse ("Muse who?"). The hype died down bc nobody could touch the models and meta/muse came out with a similarly capable model in the meanwhile.
At the same time, OpenAI had a disastrous dev day that led a lot of users to jump ship to Claude.
Couple weeks later, Gemini-4-argon is still unavailable on anything less than the ultra tier, which means there's no natural funnel to get ppl to pay for the more expensive ultra plan other than "trust us the model is good".
If Google had argon broadly available earlier this month they would have picked up a ton of new users.
As someone internal at Google I don't understand why there's so much criticism of Google models... but right now I'm using the next model beyond Argon, and Flash 3.8. In my personal Gemini app I only have access to 3.5.
Models move so fast now and Google is very cautious about releasing them.
Because Google models are crap. They are on par with local shit models. Last time I tried they were unable to even hold a longer conversation, because they would enter loop and reiterate last answer regardless of the prompt etc.
Then basic mistakes (inferring wrong information from the given text, hallucinations...). Gemini is okay for some absolute basics and the rest of the work goes to actual frontier models.
All the power users are always jumping around by now and many hold the smallest subscription with all the models or at least two.
Narrating this as if it’s the game of thrones is a fun exercise, but it’s just becoming commonplace now that models jump ahead of each other on a weekly to Monthly basis.
> and many hold the smallest subscription with all the models or at least two
This was part of the issue, as stated in their comment. We couldn't jump around to try it for the reason that is wasn't made available to, as you say, "all the power users".
I'm still a codex man myself but lots of people feel like opus 5.5 just gets it for some projects in a way that astra doesn't. Especially given the fact that astra and sol 6.1 are often at capacity late at night during US hours, when lots of people want to run longer unsupervised projects while they sleep. There's a general sense I've felt reaching some comment threads that openai failed to react with something better than opus 5.5 and are rushing to release small tweaks to make people feel like they're still getting something new.
Sol is just hilariously slow. Even Luna is slow. It is unusable compared to Claude. I was pretty unhappy with Opus 5 and really gave OpenAI a shot for personal projects, but their quotas are just too small and their capacity issues in the last few weeks have just been untenable. No good. Opus 5.5 is just a very good model, very efficient with usage, gets things done at a reasonable speed. Sonnet 5.5 is also a very nice worker improvement. It is probably on par with Opus 4.6-4.8.
Gemini 4 is announced be available to Pro users in the Gemini app.
I mean, I heard rumors too that it would be ultra only, but that doesn't seem to be the case, unless you are specifically talking about antigravity?
Realistically, market adoption isn't going to be the result of terminally online people (like me, I confess) following things way too closely on social.
There weren't "rumors" that it would be Ultra only, that's what Google has announced two weeks ago. Happy to see they've went back on that, but I haven't seen Gemini 4 on any apps till now, and I'm a Google AI Pro subscriber.
edit: just a few hours after this comment I got access to Gemini 4 on the Gemini app. Yayy! :)
> The model launches today at an introductory price of $2 per million input tokens and $10 per million output tokens (with cached input tokens at 95% off).
Couldn't you just say what the price will be after the introductory period and when it'll take effect? Do they have some SKILL.md for these blog posts that requires the most important information to always be buried ten levels deep and hidden behind a login page?
With the pace they're delivering models, there's gonna be 1-2 new Gemini 4 versions by December 1st, presumably with their own introductory price windows. We just got Gemini 3.8 on 2 September and Gemini 3.7 on 13 August.
This reminds me of telephone commercials that ran back around 1987, saying they were offering long distance calls for 10 cents a minute (a pretty good price back then) and -- this was emphasized, a huge selling point -- they guaranteed your rate would never go up. In retrospect, all the lols :-)
I wonder if that’s more to price in the cost of keeping old models around on expensive inference hardware. It’s not like old models get cheaper to run, and a more heterogeneous fleet is less efficient to manage.
> Couldn't you just say what the price will be after the introductory period
+1. FWIW, I didn't even get far enough to wonder about the price.
The lead (and only) sentence states "Our next era of frontier intelligence, built to solve your hardest problems" a statement equally true of any new Google or competitor model, hence completely devoid of useful meaning. I then kept scrolling down... and down... and down, giving up after a half dozen screens of big images, huge videos and some customer stories. The first sentence of every text block was equally vague or meaningless so... tv;dr (Too vague, didn't read).
> They don’t really. Most normies which are the vast majority of people just use ChatGPT. All tech folk just use Claude.
It was the case a year ago, but it's less clear cut nowadays, I've seen non-tech people using claude on their phone in the train, and many tech people seem to have switched from Claude Code + Opus to Codex + Sol over the past few month.
I just created a Google account, which was a hassle. I tried to find an app and found AI Studio, but I’m not sure what that is. Then I found Antigravity, which apparently is Google’s alternative to Codex, and I installed it. Then I tried to buy the small Gemini AI Ultra package, but it didn’t let me pay. It accepted neither my credit card nor my PayPal account. It’s crazy how bad the Google experience is nowadays. You don’t know what to use or where to find it, and the security measures seem totally overkill. Am I really the only one having these experiences? Is Google aware of that?
I bought a google play card (or whatever it's called) with cash due to these issues. Result: Instantly rejected + locked after entering the code, no credit, no dispute, no customer service, no refund.
Same thing, I've always paid my bills, I have had some sort of google subscription for the last 10 years. Now they're requesting me to switch to pre-paid billing. They won't accept my card (that has worked every time this year so far).
Is there anybody in Google having a look at how bad and complicated their payment system is? In all the other AI providers you just sign in, pay, get a key, done. In other clouds like AWS, you add a card, pay little to verify the card the first time, monthly bill, done. In Google it looks like you have to fight them every time for the very thing they should make as easy as possible: opening your wallet.
First install Antigravity and see if it works on Free with your account at all. You can buy Ultra and then find out you don't have "Antigravity eligibility" on a given account or it will ask you to send SMS to some nonexistent number.
I would also advise against buying Gemini if you have any other Google services, because you get banned across board if they don't like something. To use Google it's best to first migrate off of Google.
Otherwise great product, I really like and would recommend Gemini. The usage is also very generous, at least Flash models.
Using google services has always been miserable. It used to be way worse. Now you can at least get an api key from AI studio without going through 20 esoteric steps to configure everything.
I hate everything with Google and consumer facing services, they dont give a shit because 99.9% are monetized through spyware or are buisness customers.
They had gemini cli, what happened to that? Was fine
Why do these benchmarks often show models running on Max i.e. Opus 5.5 (max) is on the frontier? I would think it is never really economical to run max
In particular, where cost is the x axis, a well-funded company with a good but not best model can discount their new model until it reaches the frontier. In some sense this is as much a statement about the efficiency of markets in the presence of public measurement-based comparison and public list pricing as it is about research pace.
Interesting! Do you have webgl disabled for incognito or for chrome in general? I wonder if it's trying to do software rendering.
FWIW, I have literally the exact same configuration (M3 air, google chrome). Here's a recording on mine where I'm seeing 60fps: https://non.io/video/video-3587018401.mp4
Mine doesn't stutter on on Firefox on M1 Max, but it pins 7 cores the entire time the page is open, not just while loading. It also uses 50% of GPU in addition to that.
Edit: The argon built site pins 1.3 cpu cores. The Sol 6.1 site uses 0.18 cores.
Truly S-tier. It's sharp, creative, intuitive. Can't believe how much better it is than Opus 5, then they went ahead and made it cheaper and faster too.
I'm constantly blown away by it. I do think Astra is slightly more intelligent (although worse at this particular benchmark), but Opus 5.5 has become my default model for 99% of what I do. It's so good.
NGL, I think this might be my favorite benchmark, ha. While the Chinese labs asses are a bit off, I certainly have big hopes for their next versions. Their hearts are certainly in the right place.
Beyond SVG, these things have been RL'd on 3d reconstruction (which carries over), and I expect given the public reaction to recent frontier advancements in 3d capabilities that this will get allocated a larger share of the post training budget in the future.
What I'd kill for would be an app where you take a pic of scribbles and drawings on a whiteboard/notebook and the app turns it into SVGs you can then continue to manipulate digitally on PC/tablet.
They can represent something in vector, but it's always very hard to use the result for anything else. Like, the paths doesn't make sense, aren't grouped to the logic elements etc.
Last time I tried, it did not. That's why there's a place on the market for an app just tuned to handle the transition seamlessly, rather than having you fucking around with prompts in LLMs.
I for one don't want a whole app for that, not anymore. It's a small task whose result is an input to something else - ideal for a tool for LLM (or script) to use.
> How many pelican riding bicycle SVGs were there before this test existed? What if the training data is being polluted with all these wonky results...
But Simon has demonstrated before that asking for any animal doing something silly will still get similar results from the same model.
It's still clearly not a part of any model's training set in the sense of pelicanmaxxing in particular. It's either a capability to generate arbitrary SVGs (or even something broader), or no specific training at all.
Certainly not because all labs are pelicanmaxxing on the same exact image. Most patterns are similar everywhere. That's the result of training on the same datasets and scraped web contaminated with other models' outputs, combined with similar training methods and resulting mode collapse. Try if e.g. dolphins knitting socks in spacesuits are similar on all/most models.
Certainly true but I think the question is “does targeting the SVG pelican mean that the quality of all SVG output improves?” If so it’s a worthwhile thing to target.
It's a little bit like any benchmark, especially microbenchmarks. It is useful as far as it reflects the broader use cases, but the risk of overfitting to the specific task and not the broader ones is very high.
Sadly, this is all that managers know to do: take a metric and try to optimize the heck out of it. They don't care about research, technology, "don't be evil", etc. And Google is run by managers now.
Expect Google models to be benchmaxxed as much as possible, but don't expect commensurate performance on real-world tasks.
While this might be overly cynical, it's also a powerful validation of the exact feedback loop that you want to have in an organization.
If what you say is true, then the next set of benchmarks will be written to illuminate the true reflection of real-world tasks. Then those get benchmaxxed so managers can show "numbers go up."
Repeat enough times, and you get materially and genuinely improved products.
The burden is on the press, outside researchers, and general tech industry to both write and care about the metrics that matter.
As a cyclist, I tend to pay more attention to the bike shape than most people. Everyone knows what a bike looks like, but as with LLMs, most people would struggle to draw one - correctly connecting all the different tubes at correct angles etc.
I note that even these bikes are not actually rideable. For a bike to be stable, the front wheel axle needs to extend forward a little ahead of the fork - the fork must ”bend forwards”. This is what actually keeps the bike upright, not gyroscopic effects or some such. Given LLMs are generally good at basic maths, im surprised they get this wrong.
Maybe a bit less so in this case though. As per the announcements, the model was actually trained for this prompt, so was fed some training data. I can well imagine the data there was just not so good.
I feel like perhaps this is a bit too nuanced of a critique. While I do see what you are saying on many images of bicycles, there are also many that that I do not see it on. Close to half of them, in a quick Google image search, have forks appear straight and the axle is dead center on the ends of them. If there is something there, I am not seeing it even when trying to specifically look for it.
I'll also note that, using a piece of paper as a straight edge against my screen, there is a very slight forward curve on the low and a bit more on the high responses. Medium does appear to be perfectly straight.
Same idea in 3D: Argon on the same Three.js prompts as the rest of the field, judged by blind votes. First 4 scenes are up at about 44 cents each on high: https://threejseval.com/models/gemini-4-argon-high
Is it me or is there maybe some carryover happening where the pelicans are portrayed as expending different degrees of effort too? Like, the high effort pelican has a determined look on his face.
Is the prompt context telling the models what their effort level is, and the ambient presence of that concept of ‘high effort’ in the context window is leaking through and affecting the portrayal?
xhigh: "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer."
low: "Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration."
Recently, I’ve discovered that Claude can render great SVG versions of user interface screenshots. I ask it to substitute simple SVG friendly landscapes for any photographic images that happen to be part of the screenshots. The results are much better than they should be, especially with Opus. Usually good enough to be used on a website as a way to get an infinitely scalable, much smaller version of a screenshot. But clearly, I should try the Gemini models on this real world task. That’s where we’ll find out if all of this training to beat the benchmark actually pays off.
How many cicles before pelicans become photoreallistic SVGs and them come back to stylistic artistic abstract renderings?
Also, that first tweet contains a commentary that, for me, explain a lot about the public misconception about how current AI works:
> wait, does this mean you guys are actively training svg generation? i was mostly amazed by it thinking the models doing better at it by itself due to becoming better at reasoning.
No, it's not better reasoning that drives current AI, it's better examples we have been giving them. Whole army of people generating content for the frontier labs.
Also, reasoning improving the out of distribution part, like computer use, tool use etc.
Another case where allowing oneself to anthropomorphise LLMs a little yields better intuitions.
People don't get more skilled by raw reasoning, either. They improve their skills with better textbooks (pre-training data), more practice exercises (RL data), and generally goofing around (self-play).
Why are all these AI's making the pelican's pouch so large though? Unless they have some huge fish in there or something they don't tend to look like that... Is it that it's drawing a "cartoon" pelican or something?
It's obvious OAI hillclimbs the image generation -> SVG loop. I'd put money down on that being why OAI's pelicans have a yellowish tint that matches the yellow tint of OAI's earlier image models.
At some point Simon needs to move from Pelicans to something more complex. Every time a new model is launched by major Frontier company, I already know with certainty that first HN comment on the news would be about how Model is drawing the Pelicans.
Front bicycle wheel in all Argon examples is still wrong. This bicycle will not be stable. Axis of rotation of front wheel needs to be in-front of axis of rotation of steering wheel. Frontier VLMs still do not understand how bicycles work.
simonw, your pelicans are so famous now that the frontier labs can just RLVR for SVG images explicitly.
It's a simple manner of prompt...generate concept art with nanobanana, measure, llm creates svg with text, measure the distance between rasterized svg embedding to nanobanana art....do post-training RL on a 3 million such images and matching embeddings.
I swear this will be a canonical example of Goodhardt's Law in action.
You have to switch to a new benchmark that isn't so easy to run a verifiable rewards pipeline on...Perhaps a longer horizon task where it's easy to mess up midway.
Why? Don’t you think the steady improvement in pelican bicycling SVG art is itself a worthy field of endeavor for LLMs to be optimized for?
Yes, Simon has made pelicans a benchmark and now models have grown very good at rendering them. Is there a reason to believe that this isn’t manifesting as useful skill in general in SVG generation?
It starts to be so good that the model doesn't really matter. Each model starts to have its own personality, and it's the personality of the drawing, like a real artist, that defines the style of the pelican, and that makes it harder and harder to decide if it's good or bad.
The models that are really good at 3D modeling tend to also be really good at pelicans. I think the skills are strongly related - it's about having good 2D (and now 3D) spatial awareness.
The test is always going to be hard though, because it's an absurd image to ask for.
"Pelican on bicycle in SVG" is an adversarial example on multiple levels - out of distribution, ill choose of tool for the task, and anatomical impossibility of real pelicans using typical bycicles.
Yeah, that's a good point. My basic pelican test doesn't allow the model to render their work and iterate on it. The 3D models all come from agent systems that are rendering with Blender or screenshotting with Playwright along the way.
I’d really like to see what each model comes up with if given a basic harness to render and iterate. They’ve all gotten pretty good at basic one shot pelicans, but seeing what they do when given a direct goal, and even getting a time lapse of each image as they edit and adjust would be really fun.
A lot depends on what the models were specifically trained for using reinforcement learning. Modern models are generally OK and everything, but they're only really good at the things they have extensive RL for.
Honestly surprising if other labs aren’t training for svg generation. Feel like SVG’s are going to be really useful for dynamic UI generation stuff that goes beyond cobbling Html together
I've been dismissive of the SVG-of-pelicans benchmark, but last night ChatGPT (on the website, toggle set to GPT-6 High) generated some SVGs and included them in our chat. I'm not sure that this feature was that valuable in this case, but it seems that textually generating SVGs is now a product feature at least for ChatGPT. The SVGs had some clear issues and I found myself wishing that it had done a better job: the nose of the rocket is overlapping with the rocket body, and in the second image I think there's only supposed to be one orange line but it tried to add styling or something that went wrong.
- I was having trouble embedding the SVGs as links, and ChatGPT recommended Simon Willison's SVG Renderer as a workaround.
Just my 2 cents on the pelicans benchmark: I think it might help if you were a bit more explicit about not taking this one benchmark too seriously. I know you say that at times, but at other times (such as the comment I'm replying to) you elide it which can read as a seriousness about the benchmark.
Buddy, there's a Pelican riding it. No real cloud is like that, no real sea is like that, no real gulls are like that. And that's fine because it's a cartoon.
what is even the point of these now? labs are most definitely training on this question now. I feel like this test scratches the surface of a "real" test for these new models
It attracts at least one comment like yours on every single thread, so I guess the main purpose is to kick off a race to see who can post that comment first.
People who are unable to comment on the underlying technology or capabilities because they don't know or don't care to inform themselves sure can feel good that they have an opinion on pelican art.
The roof is always raising month over month. If you were to make a harder version of the benchmark, what do you think you'd like to see the agents do better?
Medium and high are fixies apparently, (or they have a worm gear, although the hub doesn't look big enough for that), since neither has a derailleur. That would also explain the lack of brakes at least.
Low is still obviously failing the "test". It's pedals are not actually attached to the bike either.
Medium has a handlebar that goes through the pelican's beak. So I'd say also failing.
For the high, the only nitpick would be that the bell isn't quite on the handlebars and the stripes on his scarf don't line up correctly.
The models are solving Millennium Prize questions and making progress on the Riemann Hypothesis, yet the top comment is still looking at pelicans somehow. To me, it feels even more superficial and pointless than arguing about the presidential candidates' tie choices in a debate.
It's a great model. I have had it trying to decrypt the Voynich manuscript for the past couple of days without being biased by other research/other people's attempts. If anyone is interested: https://github.com/d2kx/voynich-research
> An interactive companion to our ongoing exploration
Eyes glazed over again.
Did you even read your own site before you posted it?
One of the first pages I opened:
> Smoking-gun proof of plant-label ingestion: in Row 2, the central cut root is labeled okol (f88r.15), and directly underneath it on Line 2 (f88r.19) the recipe writes...chokol daiin qoekol qoekol qockhol okol cheol...!
How about you post something you've done/built instead of just insulting other people. Spend your mornings using people stuff just to poke holes in it?
Model announcement pages are the absolute worst. This has to be some kind of inside joke among labs, they're trying to see who can build the shittiest page.
The "Explore the latest" section is nearly blank for me - are there supposed to be diagrams or visual aids, or is it one headline and sentence taking up the whole page on cards?
Gemini isn't perfect but I find it a lot less annoying than other model families (personal preference or stockholm syndrome). Interested to read the reviews.
Seconded. Anthropic feel the need to limit how abusive you can be to Claude because god, I want to throttle that snotty little shit. Google don't have to do that with Gemini because it just isn't as irritating. It does what I ask and doesn't try to second-guess my requests, or do "the quick fix".
It's been really bad at lyrics but gotten better. You look something up and it will just make something up and it takes a bit of digging to discover it's because of copyright
Of all models I like to write with Gemini the most. The language appears much more natural and elegant than what even Opus 5.5 or Astra do. Also, they give 5TB of cloud storage and discounts, it’s a steal for what you get.
Claude has particularly bad prose. Somewhat improved in 5.5, but it has historically clearly been the most annoying prose among major models because is had so many obvious tics that made it clear it was Claude rambling. Still extremely verbose.
They all ramble, they're all allergic to getting to the point. But, Gemini historically has pretty good prose. Even the self-hosted Gemma models have pretty nice, inoffensive prose. I wouldn't want to read a novel written by any of them, but I tend to ask Gemini questions in chat more often than Claude or GPT, because it has good prose and is good at using the web to research.
Agreed. It's much better at working with stuff (like documents and presentations) in other languages too. It feels like it's trained in multiple languages from the ground up rather than as an afterthought.
I ran it through my own coding eval today using pi/OpenRouter (11 small bug fixes from my own projects). It solved 10 of 11. Compared to Sonnet 5.5 and GPT-6.1 Sol, a median run cost me $0.52 vs $0.18 for Sonnet and $0.09 for Sol, and took ~4 minutes vs 32 seconds and ~1.5 minutes.
It thinks a ton, ~24k output tokens a run vs ~3k for Sonnet. It missed a solution by thinking its way into a 10 min timeoout.
Something to look out for - if you're using pi, it doesn't have Argon in its catalog yet and reports cost at Kimi's rates, about a third of what OpenRouter actually bills.
Good God, man. Google has the worst frontier lab rollout of them all. And they're delivering products that are behind the curve constantly.
Now, I happen to believe that's an excellent strategy. I bet they actually resolve more LLM requests than any other company at this point. The lay person still uses Google exclusively.
this is why antigravity is a doomed product at a pro level.
they gave away so much pro that they need to have smaller usage levels than competitors or they'll lose a ton of money. $20 was already a loss for users actually using their quota. the majority of pro users are paying far less than that.
so they'll lose their paying users, and keep their free or subsudized, which will make the economics even worse.
just cancelled AI pro myself, since I exclusively used it for AI and also pay for youtube premium family it was becoming quite the monthly just for google
if they throw in youtube premium in that package then I might reconisder
Google One - Pro tier. I have it via Lenny's Product Pass which, imo, is a much better deal since it includes Cursor Pro, Granola, and like dozens of other paid sub tiers: https://www.lennysproductpass.com/
I use the hell out of both. Kind of sad that they took Spark out of beta-access, being able to orchestrate via your phone on something that has direct access to your Drive meant I was authoring skills and other functionality on a schedule with spreadsheets serving as databases, deterministic tests, etc.
Remembering the horror stories about lost Google accounts, I would be comfortable paying the next lower tier (100 GB) because I can backup everything myself, but not the 5TB plan, I'm sure I'll lose something if the account is lost. Hopefully these plans guarantee a human in the tech support loop.
Yeah, very true. I need to set up a recurring backup-to-external-HD automation for it. It's vanishingly easy with the app and I don't doubt they're training on all the data itself.
Google AI pro in the USA, if you have Verizon wireless it’s $10 a month add on. Without Verizon it’s $20 a month I believe. You can also share most of the benefits with up to 5? People.
Afaict, they are targeting difference user bases, Google being after their existing user base, having lost in the developer ecosystem to the two labs and open weight models.
Corporate companies. Bog-standard office workers don't need top-of-the-line Mythos intelligence, just something to add to Google Workspace. I personally doubt developers use Gemini for coding beyond marketing materials.
Are you suggesting it’s BETTER than their competitors?
Because if it’s not, and it’s not available to all paying users (Right now, it’s only available to Ultra users), then they are behind the curve if they’re treating it like Fable.
Maybe? But it's so many orders of magnitude apart. I can easily do 1B tokens a day (output tokens is probably like 3M?).
When I see non technical people use that Google reply I'd be surprised if it's more than 1k output tokens, and with a tiny model. So a 3000x in output tokens, on a model at least 10x as expensive, I wouldn't be surprised if hardware usage is 30.000x for a dev that uses coding subs vs someone who uses the Google AI overview a couple of times a day.
And that's not counting the absurd difference in cache reads.
Moreover, you need a Google account in order to use it. I wouldn't bother paying 200 USD for it but I just don't want an entire google tracking package just for using one product.
Oh, I didn't notice that they switched from 1.5 to 3 flash about six months ago. I know someone there and they talked about how 1.5 was hardened and it was a pain to do so on newer models that is wasn't worth it for a long time. But that convo was in the beginning of the year. I should have looked it up.
The model used is irrelevant when there's the critical issue where their "AI" assistant can't understand the word "nearest". Literally worse than a second-grade understanding of English.
I also have had shockingly bad results with gemini on android. Ended up giving up on it after a few months. Silly silly google with their trademark UX polish and attention to usability.
This reflects pretty much every result of mine from Google these days. Not just navigation, but search too.
The only thing I can figure is that Google's consumer-facing free-to-use models are set to "reason as fast as possible with only the most basic plausibility checks." I have to correct them like this every single time.
this is like that tiktok guy who asks the spoken ChatGPT model very simple questions, like 'count down from 60', or 'how many e's in seventeen' or whatever and the model shits the bed every time it's hilarious
I think these companies have good models but expose them poorly like others have mentioned
Often it seems like the problem with the Gemini apps is the harness, not the model itself. For example if I upload an image PDF it recognizes text worse than if I use the API.
Is this from you car ? Gemini in car is probably less intelligent than a spider.
I tell it to "Play Let it go from frozen". It plays it on youtube music. 5 seconds later I tell it again 'Play let it go from frozen" it then just dies saying "I can't play videos in car".
The auto version is extra horrible because what it reads to you is just a summary of what the model actually said. I don't ask it anything when driving because it's too frustrating.
Yeah, whatever they have running on Google Maps is probably a pretty low tier model, but it is total trash. I’ve only used it twice though and that’s when it first came out. I don’t know if there have been any changes since then. The main problem is I felt like it was not keeping track of the context of the session.
which is more or less the crux of the problem android auto's gemini implementation. It sometimes does not even know that it is in a car. The ghome implementation is also like this.
Adding my five cents: initial usage through OpenRouter has been great. Very fun model to spar with, clever, intelligent. On par with Opus 5.5 but without the Claudism's and more of a Sol type of conversation. Though speaking of Sol, I've been very dissapointed with Sol 6.1 to the extent that I feel OpenAI has neutered it compared to a few months ago. Especially when giving it goals it will happily run in dumb circles just burning tokens without any real work.
To everyone saying not available yet: it looks from the page like Pro and Ultra subscribers have it today and given the play links, it'll probably show up in the API in a few hours.
Available in agy for me as of now, I'm running my new model benmark of having it do comprehensive code audit on one of my more complex projects, it spun up 1 subagent per audit section and it took ~10 minutes per subagent on an 800k LOC, 60k complexity, ~4k file project. Will take me much longer to go through and determine the quality of the audit work.
Antigravity CLI 1.3.4
[My account email] (Google AI Ultra)
Gemini 4 Argon (High)
~/[working folder]
> /model
⎿ Model set to Gemini 4 Argon (High)
Same. It just told me "agent execution terminated due to an error". There was a little chevron underneath it that let me expand the error to view more. I was hoping to see what the error was, so I clicked it, and it said... "agent execution terminated due to an error".
It's crazy to me how care there is in most of these vibe-coded vibe-coding harnesses. I don't like oai's models much but I have to say their UI has by far the best attention to detail of any major player.
I'm on Pro and it's trying to get me to upgrade to Ultra in the AI Studio UI when using Argon. In Antigravity CLI and the Gemini website it's not showing up at all despite being logged in
I like this phrasing and am curious what new word or phrase we might use to describe big companies over pushing Ai internally at the expense of their product quality. This pattern is so ubiquitous across google cloud products I have taken to calling them the 21st century IBM, having been a IBM baby and witnessing first hand their fall from exceptionalism.
The internet and gaming has ruined my brain because scuffed was a way to describe a bad dungeon run in mythic plus dungeons in WoW, or a bad tower dive in LoL a few years ago. I see scuffed and think "here's a WoW/League knower".
HP is the other company to consider as analogous. And maybe more explicitly so, as they had a similar "engineering first" internal culture that degraded markedly over time to be a shadow of their former self.
And took similar companies that were quality engineering focused -- like DEC or Palm -- down with them through acquisition.
I couldn't see it either, then I clicked the 'Antigravity' button in the top right, switched back to the main screen, and it appeared in the model selector.
"Argon autonomously turns raw video to a physics-accurate simulation, estimating depth and physical boundaries, and writing the exact simulation code required to replicate those real-world environments."
Geometry could be understood today with the model but what also really matters is the sensing, friction, mass how smooth the subjects are, dexterity is a very difficult problem so solve, not sure how simulations could handle such factors
Game engines have many of these features. At one point, iirc, Unity was positioning to be the simulation environment for world model agents before we spoke about it in these words.
When you are on "gemini.google.com" there is a little lock icon next to Gemini 3 and Gemini 4. When you click it you get the prompt to subscribe (pay). Once you have successfully done so, the Gemini 4 option disappears!
As I approach a transitional period in my career, I'm reflecting on the managerial effort of places like google. How on earth do they keep an engineering org on the 'right track'? I work on products that are not as complex: fewer moving parts, fewer egos, no where near as novel. It feels like nobody does anything and nobody has any expertise
Makes me wonder if more growth/innovation could've happened (outside of google) if they didn't employ this many people. Some sort of dutch disease within the higher echelons of the tech sector.
Google did a very good job on things that required longer research times and compute scale, note that Transformers, BERT & Scaling Laws came out of Google (though the latter were also worked on at OpenAI).
Conversely, Google is horrible at product, and without OpenAI, it probably would have taken many years longer to take off. I know this because Google had demos of this internally, but absolutely no stomach to release it, and they didn't believe in the API use case at all.
Which is to say, I think if Google was smaller, we probably would have gotten more SaaS, not more breakthrough research.
It’s like the USA. How do they keep people on track? They don’t. You delegate and hope things work. When they have Google Ads and Search they can experiment a lot and “waste” a lot in search.
They haven't solved management and definitely don't keep things "on track" globally. However, they definitely don't have the problem that nobody does anything and nobody has any expertise.
It's also worth noting, Gemini 3.5 Pro never came out.
Having worked there (but only in a small part of it that eventually got absorbed into Gemini), it's some combination of:
* The eye of Sauron (management) can make specific spots (e.g. Gemini) push harder (and attract the motivated people at the company and be selective)
* They hire a lot better than average (Generally from the top 10%, hotter places in Google are more selective)
* They have a promo process that gives career motivated people a way to prove themselves (not always aligned with things being on the right track ofc, and not always non-political, but they definitely do something)
* Redundancy in teams causes them to compete (not necessarily productively)
Of course, there are other downsides not listed here (insane bureaucracy for launches, GDPR/BCID compliance, etc, etc), but a lot of what makes Google possible is the insane amounts of cash Search throws off.
Success for google is defined in a different way than it would be for an ordinary company and many of those would have been successful if they had started outside of google.
The really annoying ones are the ones they buy, and then shutter because they can't make enough money on them fast enough.
I have the Free AI Pro plan from when I purchased my Pixel 11 Pro. I still don't have Argon in AGY nor web nor Gemini app. I have a free product and still feel cheated.
I'm open to Google winning me over again, but there isn't enough hype in the world to do it. They will need to earn it after these disappointing 6 months.
I appreciate how it's being integrated into the Google ecosystem, but the user experience still feels unpolished when actually using the system. The marketing, however, is impressive.
I seriously doubt their ability to double API prices on Dec 1. Token prices are in a clear downward trend and Google should maximize revenue by competing for volume. Currently OpenAI is the only US frontier lab trying to seriously compete with open-weight models that you can run yourself for an order of magnitude lower than what frontier labs want to charge
Because you have an android phone and don’t want to mess around with anything else(i.e the same reason people just use internet explorer on windows) or you hold Google stock so you want to support your investment :)
Really? Reminds me of Gemini 3. It took so long before being able to get a subscription for it, and then I would run out of their equivalent of the 5-hour usage window in a heartbeat.
Yeah maybe I've just become used to how much I can use Flash 3.7/3.8 but on the less expensive Ultra plan I basically never exhaust my quota even when working over multiple projects, but this was pretty much immediate. I did ask it to audit a project but still, I do that in Flash regularly and it's not especially costly.
I made a NHL inspired hockey web game with Opus 5.5, and I let Argon loose on it to try and see if it could improve it, and it found a ton of bugs and improved performance and load times significantly. If anyone wants to play it here:
I would love to try it out, and the Gemini models had previously been some of the ones I'd liked best in an overall sense, but unfortunately their updated inference endpoint is no longer ZDR on OpenRouter (change takes effect on the 18th). I think this is one of the more unfortunate turns in AI development in 2026, as it seemed like labs offering it was becoming the norm last year, and Google had previously done so consistently.
AI Studio has never been ZDR (retention is noted at 55 days max), but paid usage is excluded from training. Vertex uses enterprise terms, and it's definitely ZDR/no training. What's happening on the 18th?
Only available on Ultra, so I thought maybe I'll try it for one month to see if it's good. Well, I have Google One and for some reason this means I cannot sign up for AI Ultra? This is just one of the worst product rollouts I've ever seen
I checked gemini, set it to "gemini 4 argon" and it offered me Pro and Ultra to get access. I checked the normal plans, saw there is also plus.
Okay, so I'll get Pro to try it.
I subscribe.
No gemini 4.
The most generous reading I have is that it's a mistake. But I also downloaded the PC app, and now I really wonder why I did that, it seems to offer nothing beyond the web experience and also doesn't even know its a local app. (When asked about its capabilities it always responded with reasons pertaining to it being a web app.)
Right now there are two top posts one about GLM-Legal and one about Argon. Both score about the same in the Harvey Lab test, but the prices are $0.15 versus $10.55, the $0.15 is Vals AI's estimate using the base model since GLM-Legal has not published prices.
Yeah, but one is a general purpose model, and the other one is specifically geared towards the legal field. Considering Opus 5.5 only gets a 4% on the test compared to Gemini’s 20%, that’s actually pretty incredible.
I gave the model my usual question of "I got a raise from X to Y, please explain what is the impact on my net salary". Answer was detailed and nailed all payroll edge cases, only to miserably fail in basic math at the end. It tried to calculate 430 minus 60 and got 410.
Is there a benchmark for reverse engineering compiled programs?
- Grader compiles a program to binary -> Agent has to reverse it back, and match all bugs and behavior.
Digging in the model card, "Gemini 4 is the next generation in the Gemini series, a suite of highly-capable, natively multimodal, reasoning models. Gemini 4 Argon delivers frontier performance..." If they're using noble gases there are 2 obvious lighter naming candidates and 3 larger (or 4 if you count Oganesson). I'm not seeing them claiming Argon is the top/biggest/bestest model among the Gemini 4 series, just that it has frontier performance.
Edit: noble gases not noble taxes, though the latter would be a funnier naming scheme.
I hope all the people very concerned about AI safety are cheering this one. The best guard against one AI swarm taking over is having multiple adversarial orgs all with similarly advanced AI swarms.
I wonder if any of the AI safety people will begin encouraging AI researchers to swap jobs to work at a failing lab in order to keep them all extremely competitive with one another.
I'm still waiting for a model which I can tell to "take these 3 screenshots and make a product launch video", where it doesn't randomly modify the screenshots or garble/rewrite the text as it moves the screenshots around the video.
Every time I try to generate one it's like a slot machine, where a different part will be messed up after fixing the first.
I dont have that problem with opus5.5. I dont even take the screenshots anymore (it does all screen recordings)... I tell it a simple narrative and it follows through. e.g. https://x.com/plannotator/status/2108589088139886769?s=20
> Across Google and our early access partners, Argon is quickly becoming the preferred model for engineers tackling long-horizon software engineering.
"Across Google", wonder how much that is because of an effort by Google to curb usage of Anthropic models. MSFT and Meta similarly are pushing for in-house models. It doesn't mean the models are better when you push your engineers to dogfood and restrict access to other LLMs.
(Not that I know for sure this is the case, it's more of a hypothesis).
It passed the test of using latest crystal-lang concurrent feature (with web search) when gemini 3.8 flash failed (again, with web search enabled). I think it is at least useful for coding, for sometime.
Pricing is interesting. At launch Argon is the same price as GPT-6 Sol, half the price of Claude Opus 5.5 and 1/5th the price of Claude Fable 5.1 and GPT-6 Astra.
But on December 1st Argon doubles in price - so it will be the same price as Opus 5.5.
I've kinda stopped looking at token pricing. Argon might be the same token pricing as Sol, but in Artificial Analysis' testing, Argon costs 5.8x more to run their coding tasks and 2.2x more for their general index.
Likewise, you look at Artificial Analysis' general index and it cost $8,708 to run with Opus and $2,407 to run with Argon so even if Argon's token price doubles, it won't be the same price as Opus.
Sonnet is cheaper than Opus per-token, but in their coding tests Sonnet burns through enough tokens that it's 7.5% more expensive.
So much depends on how token efficient something is and whether it gets you a good result or not, right?
Plus, if one is using a plan instead of paying token rates, what does my $200 get me from OpenAI, Anthropic, or Google?
It just feels very hard to compare pricing given that the number of tokens used can vary so much. In Artificial Analysis' coding tests, Sonnet used 7.1x more tokens than Sol. Sonnet and Sol might have the same per-token pricing, but there's a huge difference in the number of tokens used.
All AI pages now include canvas animations. There isn't a single page without one. This is neither good nor bad, just an observation. I think OpenAI has the best animations so far, with Anthropic in second place. This one looks a bit off.
Page says it is now available to Ultra and Pro subscribers. Although in AI Studio it currently says you need Ultra. But it's technically released either way.
Yeah, I think this is it! Google tried to hype their model but they have no chance against queen Dario. Perhaps they should just delete Gmail and say that Gemini did it. That way they would get some attention …
It seems the model is a little verbose, using more tokens to complete a task that other similar models as stated more precisely in the linked article about Artificial Analysis's work on this model.
The slides they show in the catering video demo are so bad (in addition to having all the usual signs of AI design), I would genuinely laugh if a supplier sent me this.
Finally, gemini 3.8 has been falling way behind opus and sol in my experience. It's showing up in my antigravity. My work account though still has 3.6, 3.8, 3.1.
The previous gen pro used to have this weird tick where if you ask it a reasonably open ended question, it would number things in the response and assign each one some weird label starting with 'The...'
e.g.
X: The Parameters
Y: The "System"
Z: The "Ongoing Grind"
The first two prompts I tested in AI Studio does the same, unfortunately.
Another quite unpleasant tick was that if you specify a peculiar term in your prompt as a conversational shortcut, it would keep repeating that throughout the conversation. e.g. "...What is a bitter lesson pilled way to do this?", then you'd see it will quote "bitter lesson pilled" at like 7 places in its response. This one seems to have eased off a little.
Just switched, and it seems to be much better. I'm hoping it stops lying. I have Pro max subs for Claude, Codex, Grok, and Google AI... I was forced into using Google by running out of credits elsewhere.
I have found that Gemini often claims it did something, but didn't actually complete the task.
I just had 3.8 Flash do a major refactor -- that never should have needed to be done in the first place -- and switched to Argon 4, and had it review it.
Almost everything had been stubbed out by 3.8, and none of it was properly integrated into the codebase. :facepalm:
Unlike most people it seems, I like all the progress in AI but I really do wish they would reign in the over-the-top marketing. I mean, can we just get this stuff to the point where I can reliably work with it to design a well-architected software project without having to constantly struggle against bloat and nonsense directions no experienced or sane developer would ever take.
fyi: argon is an interesting choice of name, coming from the Greek argos meaning "lazy or living w/out labor".
Glad to see them abandon the absolute train wreck of their model naming "strategy".
Gemini Live vs. Gemini Lite vs. Gemini Pro vs. Gemini Flash vs. Nano Banana Lite (that one apparently has no Gemini prefix?)... Please, just pick a metaphor and stick with it without wildly mixing reasoning modes, model sizes, and input/output modalities in the model name!
can for the love of god we get some usage in cheaper subscriptions being Ultra tier only is annoying. I want to just test it out here, well will burn some money via api I guess.
>"Argon helped the Chrome team separate core webpage handling from a ~6 million line legacy folder of browser features held within Chromium’s 40 million-line codebase. Piping multiple clean code changes at a time, Argon helped the team finish an architectural refactor across hundreds of files, completing months of work in weeks."
Now that needed to be done for some time!
Regardless of if an AI helped or not, regardless of whether that was Google's or some other AI or not (though Argon does sound pretty cool!), regardless of whether an AI cut the time from months to weeks or not, refactoring Chrome -- decomposing it into individual clearly demarcated non-tightly coupled software parts with core webpage handling at the center -- that needed to be done for some time!
Now all of that being said, Gemini 4 Argon does sound pretty cool, and we congratulate both teams at Google -- the team that created Gemini 4 Argon, and the team that refactored Chrome with it!
But all of that being said, while the advancements made for Gemini 4 Argon are important and impressive (and they are!)... refactoring Chrome properly (IMHO) is slightly more important and impressive!
(You know, one man's humble opinion! That is, I am not the guy who is responsible for awarding performance-based bonuses at Google at the end of the year, so take everything I say with the proverbial "grain of salt"! :-) I am probably biased, one way or the other! :-) )
Why can't the Gemini model code?
I'd guess Gemini probably used the best dataset.
Even Google is the place where the world's best coders are gathered, isn't it?
Yet the GPT code from OpenAI, which collaborated with Microsoft, is beautiful, while the Gemini model's coding is trash.
I have no problem with Gemini Flash 3.8 for coding, though I only use Elixir which is one of the better languages for LLMs. Where does it fall short elsewhere?
Meanwhile I paid a subscription and they only provide access to 3.6 Flash. Consequently, I have moved on to open-weight models. It is what happens when you screw your customers over.
3.1 Gemini was amazing for anything non-agentic. I switched to it for my primary chatbot purposes.
I never touch 3.8, it's awful. Curious to try this, but I've yet to use any Gemini that is good for agentic work.
Also, unrelated but this made me look at Artificial Analysis to see where it lands on Cost per Task, and it ranks Opus 5.5 at almost double Astra 6. That does not align with my experience.
Not available for AI Pro subscribers in Antigravity. What a disappointment.
Google...why. You keep getting so close but constantly fumble at the last moment. I can use Opus 5.5 and Sol/Astra on the $20 Anthropic and OpenAI plans.
You are already playing from behind. You need to GET YOUR MODELS INTO MORE PEOPLES HANDS by any means possible. Continuing to make it difficult to use your products DOES NOT HELP.
This model might be great but the majority aren't going to subscribe for $100/month just to try it.
Blatantly obvious that models will not slow down. For any task, if we can build a dataset or environment of verifications, it can be automated. This will eat all of human tasks.
The fact that multiple companies can breach any benchmark means there is no ambiguity in terms of how to build AI, its just a matter of time.
Certainly Meta and Google are benchmaxxing to please public markets, but it still proves the above
but they have slowed down. the delta in capability between model releases now is smaller than ever before and i expect it to continue to diminish until we reach some kind of asymptotic maximum, which hopefully will mean turn towards optimization and local model improvements. The current frontier models running at 10x the tok/s would be incredible for my use cases
The delta isn't really slower, we're just unlocking capability bundles in batches. The capabilities bundles you care about might have mostly been unlocked already, and just need refinement, which is coming slower than the initial capabilities came, but don't discount the breadth of advances being made across all possible tasks.
We're not living in the same universe. Look at capability year-over-year. The asymptotic maximum is wishful thinking. Someone needs to step in before this gets wildly out of control.
> Someone needs to step in before this gets wildly out of control.
This sounds like a paid bot, lmao.
I wonder how much is actual improvements in the model and how much comes from better harnesses. I'd be willing to bet that the combination of better harnesses + the models being trained around agentic workflows (i.e the harnesses) has made it significantly better in many things, but it's hard to find other improvements imo.
You can see the difference between GPT 3, 4, 5, 6 each year. Same with Gemini and Claude. Look at image and text generation. Code generation. Math ability (Navier Stokes, Quasi-Riemann.)
Year-over-year, yes I would agree with you, but that is not the cycle time of models. And while I share the same hunch as GP, I'm waiting another cycle before putting more confidence behind it.
A little over a year ago they were winning their first gold in an elite high school math competition. Have you not seen what they are doing in math now?
i burn through a $200 claude and $200 codex plan every week. opus 5.5 is not "step change" it's just a bit better, and the benchmarks show this. Honestly with my workflow I cannot really tell the difference between the top models, they're all pretty much the same in capability, just a bit different. You can build software with all of them, and with all of them you have to validate the change thoroughly to not ship stupid bugs
I do not see the current training regime of "relentlessly try anything to get the task done" as a true step change in intelligence. This is one of the reasons I suspect capability acquisition is slowing.
A more intelligent model would not default to spaghetti throwing, it would pause to consider the problem fundamentals, the accuracy of assumptions, and ask questions that remove ambiguity. This however does not lend itself to increasing token sales while prices remain inflated from market and brand dominance.
Models just solved a Millennium price and you think their progress is slowing down? I will concede that their capabilities vary a lot from one field to another, but still.
As of my knowledge cutoff, three of the top state-of-the-art large language models are Google's Gemini 2.5 Pro, OpenAI's o3 (and GPT-4.5), and Anthropic's Claude 3.7 Sonnet.
I've been trying it out via the API and the model has been a delight so far. Clear in the head, not RL-fried, distinct writing style, and big model smell. That said, the em-dash is back w/ this model.
That said, I would be so frustrated with Sundar Pichai if I was on the Google team. The botched rollout of this model seems to be entirely a matter of capacity, and Pichai has been more than happy to give it to anybody but his own team.
I know it's a matter of personal taste but it reminds me of a grad student that tries to sound smarter than he is by being dismissive and opaque.
noooooooo
At the same time, OpenAI had a disastrous dev day that led a lot of users to jump ship to Claude.
Couple weeks later, Gemini-4-argon is still unavailable on anything less than the ultra tier, which means there's no natural funnel to get ppl to pay for the more expensive ultra plan other than "trust us the model is good".
If Google had argon broadly available earlier this month they would have picked up a ton of new users.
Models move so fast now and Google is very cautious about releasing them.
Then basic mistakes (inferring wrong information from the given text, hallucinations...). Gemini is okay for some absolute basics and the rest of the work goes to actual frontier models.
Narrating this as if it’s the game of thrones is a fun exercise, but it’s just becoming commonplace now that models jump ahead of each other on a weekly to Monthly basis.
This was part of the issue, as stated in their comment. We couldn't jump around to try it for the reason that is wasn't made available to, as you say, "all the power users".
Do you really think people in google are happy about how long it took to release this model?
We’re all running on Astra medium the whole week and loving it.
I’m curious what I missed.
Can you elaborate on this?
I mean, I heard rumors too that it would be ultra only, but that doesn't seem to be the case, unless you are specifically talking about antigravity?
Realistically, market adoption isn't going to be the result of terminally online people (like me, I confess) following things way too closely on social.
edit: just a few hours after this comment I got access to Gemini 4 on the Gemini app. Yayy! :)
Oh yeah. Completely forgot about that. Big headlines and big claims, but no one could actually use it and report on it.
Does seem like they don't quite grasp how the hype train operates. Or maybe they don't care as much.
Given FelonyBench, do we also have a FantasyBench?
Couldn't you just say what the price will be after the introductory period and when it'll take effect? Do they have some SKILL.md for these blog posts that requires the most important information to always be buried ten levels deep and hidden behind a login page?
$2/million input and $10/million output today. $4/$20 after December 1st.
Is Elon Musk doing their pricing?
I think at this point comparing raw token pricing is often apples to oranges.
It is price fixing, nothing else
+1. FWIW, I didn't even get far enough to wonder about the price.
The lead (and only) sentence states "Our next era of frontier intelligence, built to solve your hardest problems" a statement equally true of any new Google or competitor model, hence completely devoid of useful meaning. I then kept scrolling down... and down... and down, giving up after a half dozen screens of big images, huge videos and some customer stories. The first sentence of every text block was equally vague or meaningless so... tv;dr (Too vague, didn't read).
Even the Gemini sub-reddit is laughing: https://www.reddit.com/r/GeminiAI/comments/1x2dr47/the_lates...
It was the case a year ago, but it's less clear cut nowadays, I've seen non-tech people using claude on their phone in the train, and many tech people seem to have switched from Claude Code + Opus to Codex + Sol over the past few month.
Is there anybody in Google having a look at how bad and complicated their payment system is? In all the other AI providers you just sign in, pay, get a key, done. In other clouds like AWS, you add a card, pay little to verify the card the first time, monthly bill, done. In Google it looks like you have to fight them every time for the very thing they should make as easy as possible: opening your wallet.
I would also advise against buying Gemini if you have any other Google services, because you get banned across board if they don't like something. To use Google it's best to first migrate off of Google.
Otherwise great product, I really like and would recommend Gemini. The usage is also very generous, at least Flash models.
They had gemini cli, what happened to that? Was fine
https://arena.ai/leaderboard/agent/chat/pareto
For almost same intelligence (+15.56%, +15.18%), drastic difference in cost/task ($3.51 vs $0.41)
Argon build: https://html.non.io/lcars-argon/
GPT 6.1 sol: https://html.non.io/lcars-gpt-6.1-sol/
Opus 5.5 (currently SOTA for this): https://html.non.io/lcars-opus-5.5/
Original designs: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...
(Chrome, also incognito)
FWIW, I have literally the exact same configuration (M3 air, google chrome). Here's a recording on mine where I'm seeing 60fps: https://non.io/video/video-3587018401.mp4
Edit: The argon built site pins 1.3 cpu cores. The Sol 6.1 site uses 0.18 cores.
Here's a vid of both opus and argon side by side, which might be more viewable on mobile: https://x.com/pwnies/status/2110058987462988152
Here's low, medium, high. Even low is better than pelicans from any other lab: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
low cost 10.185 cents, medium was 11.098 cents, high was 17.166 cents.
The costs are similar because it used more output tokens than reasoning tokens for all three pelicans, even for high.
> If it's on HN and is a meme at this point, it will end up in the training set.
> It's kind of fun to imagine that there is an intern in every AI company furiously trying to get nice looking svg pelicans on bicycles.
Google listened and assigned a few interns, I guess, for the task
Figure of speech.
>Why not just generate svgs with any editing software?
How do you do that from the whiteboard/notebook?
> How many pelican riding bicycle SVGs were there before this test existed? What if the training data is being polluted with all these wonky results...
But Simon has demonstrated before that asking for any animal doing something silly will still get similar results from the same model.
Expect Google models to be benchmaxxed as much as possible, but don't expect commensurate performance on real-world tasks.
This is not true. It is needless negative hyperbole that IMHO doesn't belong on HN.
If what you say is true, then the next set of benchmarks will be written to illuminate the true reflection of real-world tasks. Then those get benchmaxxed so managers can show "numbers go up."
Repeat enough times, and you get materially and genuinely improved products.
The burden is on the press, outside researchers, and general tech industry to both write and care about the metrics that matter.
They look remarkably good, sure -- but not perfect.
I note that even these bikes are not actually rideable. For a bike to be stable, the front wheel axle needs to extend forward a little ahead of the fork - the fork must ”bend forwards”. This is what actually keeps the bike upright, not gyroscopic effects or some such. Given LLMs are generally good at basic maths, im surprised they get this wrong.
Maybe a bit less so in this case though. As per the announcements, the model was actually trained for this prompt, so was fed some training data. I can well imagine the data there was just not so good.
People drawing bikes, scroll down to see they having made renders of all the drawings. Hilarious
You can also ride a negative rake fork. They are less stable, yes, but a far cry from not actually rideable, cf https://www.instagram.com/yourforkisonbackwards/
I'll also note that, using a piece of paper as a straight edge against my screen, there is a very slight forward curve on the low and a bit more on the high responses. Medium does appear to be perfectly straight.
https://www.walmart.com/ip/Huffy-26-inch-Rock-Creek-Men-s-Mo...
(apologies if I'm misunderstanding).
You're nit-picking on a detail no one actually cares about.
Subtle highlights on the metal frame, sweat beads, jaunty little hat…it has style. At least compared to the rather literal ChatGPT outputs. https://static.simonwillison.net/static/2026/gpt-pelicans-gr...
Is the prompt context telling the models what their effort level is, and the ambient presence of that concept of ‘high effort’ in the context window is leaking through and affecting the portrayal?
https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/chat_templ...
xhigh: "Reasoning effort is set to xhigh. Please think carefully through the task, validate key assumptions, consider plausible alternatives, and prioritize correctness, consistency, and clarity in the final answer."
low: "Reasoning effort is set to low. Keep your thinking brief and focused, moving directly to the conclusion without unnecessary elaboration."
Argon: https://jonclegg.github.io/pacman-bakeoff/dev/#gemini-4-argo...
Opus 5.5 https://jonclegg.github.io/pacman-bakeoff/dev/#claude-opus-5...
Also, that first tweet contains a commentary that, for me, explain a lot about the public misconception about how current AI works:
> wait, does this mean you guys are actively training svg generation? i was mostly amazed by it thinking the models doing better at it by itself due to becoming better at reasoning.
No, it's not better reasoning that drives current AI, it's better examples we have been giving them. Whole army of people generating content for the frontier labs.
Also, reasoning improving the out of distribution part, like computer use, tool use etc.
Another case where allowing oneself to anthropomorphise LLMs a little yields better intuitions.
People don't get more skilled by raw reasoning, either. They improve their skills with better textbooks (pre-training data), more practice exercises (RL data), and generally goofing around (self-play).
https://en.wikipedia.org/wiki/Pelican#/media/File:Pelikan_Wa...
It's a simple manner of prompt...generate concept art with nanobanana, measure, llm creates svg with text, measure the distance between rasterized svg embedding to nanobanana art....do post-training RL on a 3 million such images and matching embeddings.
I swear this will be a canonical example of Goodhardt's Law in action.
You have to switch to a new benchmark that isn't so easy to run a verifiable rewards pipeline on...Perhaps a longer horizon task where it's easy to mess up midway.
Yes, Simon has made pelicans a benchmark and now models have grown very good at rendering them. Is there a reason to believe that this isn’t manifesting as useful skill in general in SVG generation?
I've noticed that chatgpt is working harder. But still I prefer Claude, it just feels nicer to me.
As one point, the high pelican's arm is see-through/messed up.
They do look worlds better than previously, to be fair.
Modern models can make nearly photorealistic games in threejs. So why do they struggle with these pelicans?
Are you sure they're not fetching free asset packs for that?
The test is always going to be hard though, because it's an absurd image to ask for.
UPDATE: I like this theory about using an agentic harness that can render and iterate better: https://news.ycombinator.com/item?id=50073879#50074395
https://www.gianlucagimini.it/portfolio-item/velocipedia/
HN discussion: https://news.ycombinator.com/item?id=17699017
https://www.wired.com/2016/04/can-draw-bikes-memory-definite...
The pretty 3d stuff you see are agents controlling a Blender MCP in a loop where they look at the scene, think, adjust the scene, repeat.
Here's 'Generate an SVG of an armadillo in fishnet tights jaywalking on Mars' for Gemini 4 Argon: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Some other models for comparison: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
wasn't the whole point of pelicans is that labs *wouldn't* train for that?
so now the benchmark is ruined
<https://tools.simonwillison.net/svg-render#%3Csvg%20font-fam...> <https://tools.simonwillison.net/svg-render#%3Csvg%20font-fam...>
- I was having trouble embedding the SVGs as links, and ChatGPT recommended Simon Willison's SVG Renderer as a workaround.
Just my 2 cents on the pelicans benchmark: I think it might help if you were a bit more explicit about not taking this one benchmark too seriously. I know you say that at times, but at other times (such as the comment I'm replying to) you elide it which can read as a seriousness about the benchmark.
Give me reverse engineering bench and agentic coding.
Leave making visuals to humans who actually have aesthetic sense unlike whatever made this mess
Medium and high are fixies apparently, (or they have a worm gear, although the hub doesn't look big enough for that), since neither has a derailleur. That would also explain the lack of brakes at least.
Low is still obviously failing the "test". It's pedals are not actually attached to the bike either.
Medium has a handlebar that goes through the pelican's beak. So I'd say also failing.
For the high, the only nitpick would be that the bell isn't quite on the handlebars and the stripes on his scarf don't line up correctly.
Eyes glazed over again.
Did you even read your own site before you posted it?
One of the first pages I opened:
> Smoking-gun proof of plant-label ingestion: in Row 2, the central cut root is labeled okol (f88r.15), and directly underneath it on Line 2 (f88r.19) the recipe writes...chokol daiin qoekol qoekol qockhol okol cheol...!
It's just gibberish.
What kind of progress was made with Argon with the Voynich?
It looks like you've managed to decode it? Is that legit?
Like a working scroll on a webpage.
God damn this is impossible to use of safari mobile.
Argon is trying to solve the "hard" problems like those.
Is it actually available?
The code is bad though.
They all ramble, they're all allergic to getting to the point. But, Gemini historically has pretty good prose. Even the self-hosted Gemma models have pretty nice, inoffensive prose. I wouldn't want to read a novel written by any of them, but I tend to ask Gemini questions in chat more often than Claude or GPT, because it has good prose and is good at using the web to research.
Or else it say, here is the result. (but no it didn't do no research, and tells me, sorry I forgot to search).
It's got balls to lie straight at my face.
It thinks a ton, ~24k output tokens a run vs ~3k for Sonnet. It missed a solution by thinking its way into a 10 min timeoout.
Something to look out for - if you're using pi, it doesn't have Argon in its catalog yet and reports cost at Kimi's rates, about a third of what OpenRouter actually bills.
Now, I happen to believe that's an excellent strategy. I bet they actually resolve more LLM requests than any other company at this point. The lay person still uses Google exclusively.
they gave away so much pro that they need to have smaller usage levels than competitors or they'll lose a ton of money. $20 was already a loss for users actually using their quota. the majority of pro users are paying far less than that.
so they'll lose their paying users, and keep their free or subsudized, which will make the economics even worse.
if they throw in youtube premium in that package then I might reconisder
I use the hell out of both. Kind of sad that they took Spark out of beta-access, being able to orchestrate via your phone on something that has direct access to your Drive meant I was authoring skills and other functionality on a schedule with spreadsheets serving as databases, deterministic tests, etc.
Hope they roll it out to a full feature soon
More specifically, who are they targeting with Argon and why is it behind the Ultra plan? Who is paying $99/month if not developers?
Corporate companies. Bog-standard office workers don't need top-of-the-line Mythos intelligence, just something to add to Google Workspace. I personally doubt developers use Gemini for coding beyond marketing materials.
There is another subset in the Google Workspace for enterprise clients.
Because if it’s not, and it’s not available to all paying users (Right now, it’s only available to Ultra users), then they are behind the curve if they’re treating it like Fable.
It’s tone deaf.
Unique users, maybe. But a simple query that's answered in a paragraph is negligible compared to devs running agents continuously for tens of hours.
When I see non technical people use that Google reply I'd be surprised if it's more than 1k output tokens, and with a tiny model. So a 3000x in output tokens, on a model at least 10x as expensive, I wouldn't be surprised if hardware usage is 30.000x for a dev that uses coding subs vs someone who uses the Google AI overview a couple of times a day.
And that's not counting the absurd difference in cache reads.
Me: Gemini, navigate to nearest McDonalds.
Gemini: [Lists 10 McDonalds locations wnd asks which I want.]
Me: What were the priginal instructions just now?
Gemini: [Recites imperatives about tone and content from its memory system.]
Me: No. I told you to navigate to the NEAREST McDonalds. There's only one NEAREST McDonalds. Do that one.
Gemini: My mistake, navigating to the nearest McDonalds. [Opens Maps, navigating to a McDonalds across town.]
I also have had shockingly bad results with gemini on android. Ended up giving up on it after a few months. Silly silly google with their trademark UX polish and attention to usability.
Flash lite is very noticeably more idiotic than 3.8 Flash. It's why the quick AI overview is often much worse than the AI mode one.
1. I ask it to list out all my old Keep notes that are just throwaway task lists, reminders, etc. to propose deleting them.
2. It comes up with like 4 of them (I know for a fact there are dozens, if not hundreds). I ask it for an exhaustive list.
3. I see its thinking trace briefly show "Refusing to produce an exhaustive list..." and then it says "Yes, that is all of them."
Not sure whether it's lack of intelligence, bad tools, or both, but that was extremely frustrating to see.
The only thing I can figure is that Google's consumer-facing free-to-use models are set to "reason as fast as possible with only the most basic plausibility checks." I have to correct them like this every single time.
If it had exact location access people would be crying that Gemini steals their location data.
I think these companies have good models but expose them poorly like others have mentioned
I tell it to "Play Let it go from frozen". It plays it on youtube music. 5 seconds later I tell it again 'Play let it go from frozen" it then just dies saying "I can't play videos in car".
i benchmarked 20+ models in real worlds and gemini families are the best (highest customer satisfaction)
> Gemini 4 Argon (Medium)
edit: It failed on the second prompt and they disabled it now.
https://antigravity.google/blog/gemini-4-argon-in-google-ant...
> Gemini 4 Argon is available today for Google AI Pro and Ultra subscribers
EDIT:
Apparently Ultra subscribers gets Argon everywhere including in Google Antigravity
Pro subscribers only get them in Gemini app and Search
https://one.google.com/about/google-ai-plans/?icid=microsite...
It seems to be rolling out gradually.. not available at once to everyone.
It's crazy to me how care there is in most of these vibe-coded vibe-coding harnesses. I don't like oai's models much but I have to say their UI has by far the best attention to detail of any major player.
It doesn't show in the desktop UI.
I like this phrasing and am curious what new word or phrase we might use to describe big companies over pushing Ai internally at the expense of their product quality. This pattern is so ubiquitous across google cloud products I have taken to calling them the 21st century IBM, having been a IBM baby and witnessing first hand their fall from exceptionalism.
And took similar companies that were quality engineering focused -- like DEC or Palm -- down with them through acquisition.
This is in the Gemini app on macOS.
This is crazy if it can work accurately.
Pretty misleading.
is it just 'Hire the right people'?
It's extremely inconsistent internally as well.
Look up the slime mold analogy for Google.
Google did a very good job on things that required longer research times and compute scale, note that Transformers, BERT & Scaling Laws came out of Google (though the latter were also worked on at OpenAI).
Conversely, Google is horrible at product, and without OpenAI, it probably would have taken many years longer to take off. I know this because Google had demos of this internally, but absolutely no stomach to release it, and they didn't believe in the API use case at all.
Which is to say, I think if Google was smaller, we probably would have gotten more SaaS, not more breakthrough research.
It's also worth noting, Gemini 3.5 Pro never came out.
Having worked there (but only in a small part of it that eventually got absorbed into Gemini), it's some combination of:
* The eye of Sauron (management) can make specific spots (e.g. Gemini) push harder (and attract the motivated people at the company and be selective) * They hire a lot better than average (Generally from the top 10%, hotter places in Google are more selective) * They have a promo process that gives career motivated people a way to prove themselves (not always aligned with things being on the right track ofc, and not always non-political, but they definitely do something) * Redundancy in teams causes them to compete (not necessarily productively)
Of course, there are other downsides not listed here (insane bureaucracy for launches, GDPR/BCID compliance, etc, etc), but a lot of what makes Google possible is the insane amounts of cash Search throws off.
Perhaps one of their more successful chat clients, but that's all I can think of
The really annoying ones are the ones they buy, and then shutter because they can't make enough money on them fast enough.
What if the answer is "they don't"?
Source? Their documentation says it's available on the $20 plan and I can confirm firsthand.
https://one.google.com/intl/en_us/about/google-ai-plans/
https://x.com/OfficialLoganK/status/2110038940841545827
it needs MCP asap
What about the Google One AI plans? Will they get it? That has a Pro AI plan as well.
https://playblueline.pages.dev/
Even after using the free model in Google search talking to Anthropic or OpenAI models feels like talking to an ox.
Okay, so I'll get Pro to try it.
I subscribe.
No gemini 4.
The most generous reading I have is that it's a mistake. But I also downloaded the PC app, and now I really wonder why I did that, it seems to offer nothing beyond the web experience and also doesn't even know its a local app. (When asked about its capabilities it always responded with reasons pertaining to it being a web app.)
It feels so complicated and hard to know what we can use vs what hands them our codebases
On safari the page doesn't scroll very well on the first section
Unavailable · Gemini 4 Argon is currently available to Ultra plan subscribers.
Edit: noble gases not noble taxes, though the latter would be a funnier naming scheme.
I wonder if any of the AI safety people will begin encouraging AI researchers to swap jobs to work at a failing lab in order to keep them all extremely competitive with one another.
At this point a lot of stuff happens at the product level.
OpenAI and anthropic have distribution and data advantages now that are hard to replicate.
Every time I try to generate one it's like a slot machine, where a different part will be messed up after fixing the first.
I told it to strip screenshots from our old basic explainer video to generate modern 3D video.
Pretty happy with the result. I told it to keep the audio and match the video and it did.
I previously tried with Seedance 2.5, Flux 3, Kling 3, Gemini Omni without much luck.
"Across Google", wonder how much that is because of an effort by Google to curb usage of Anthropic models. MSFT and Meta similarly are pushing for in-house models. It doesn't mean the models are better when you push your engineers to dogfood and restrict access to other LLMs.
(Not that I know for sure this is the case, it's more of a hypothesis).
It still thinks Charlie Kirk is alive, though.
But on December 1st Argon doubles in price - so it will be the same price as Opus 5.5.
Likewise, you look at Artificial Analysis' general index and it cost $8,708 to run with Opus and $2,407 to run with Argon so even if Argon's token price doubles, it won't be the same price as Opus.
Sonnet is cheaper than Opus per-token, but in their coding tests Sonnet burns through enough tokens that it's 7.5% more expensive.
So much depends on how token efficient something is and whether it gets you a good result or not, right?
Plus, if one is using a plan instead of paying token rates, what does my $200 get me from OpenAI, Anthropic, or Google?
It just feels very hard to compare pricing given that the number of tokens used can vary so much. In Artificial Analysis' coding tests, Sonnet used 7.1x more tokens than Sol. Sonnet and Sol might have the same per-token pricing, but there's a huge difference in the number of tokens used.
This was hardly a consideration a year ago, when many model tasks could be tested using a single prompt and it was hard to spend more than a dollar.
Which one is it?
https://youtu.be/DlTNN0gvkLM
e.g.
The first two prompts I tested in AI Studio does the same, unfortunately.Another quite unpleasant tick was that if you specify a peculiar term in your prompt as a conversational shortcut, it would keep repeating that throughout the conversation. e.g. "...What is a bitter lesson pilled way to do this?", then you'd see it will quote "bitter lesson pilled" at like 7 places in its response. This one seems to have eased off a little.
That was a long wait.
What’s the point lol. I’m not subbing to ultra to have it
"Hans, we're doing a new point release! What isn't crossed out on the periodic table yet?"
I have found that Gemini often claims it did something, but didn't actually complete the task.
I just had 3.8 Flash do a major refactor -- that never should have needed to be done in the first place -- and switched to Argon 4, and had it review it.
Almost everything had been stubbed out by 3.8, and none of it was properly integrated into the codebase. :facepalm:
fyi: argon is an interesting choice of name, coming from the Greek argos meaning "lazy or living w/out labor".
Gemini Live vs. Gemini Lite vs. Gemini Pro vs. Gemini Flash vs. Nano Banana Lite (that one apparently has no Gemini prefix?)... Please, just pick a metaphor and stick with it without wildly mixing reasoning modes, model sizes, and input/output modalities in the model name!
Start preparing Gemma 5 release.
Now that needed to be done for some time!
Regardless of if an AI helped or not, regardless of whether that was Google's or some other AI or not (though Argon does sound pretty cool!), regardless of whether an AI cut the time from months to weeks or not, refactoring Chrome -- decomposing it into individual clearly demarcated non-tightly coupled software parts with core webpage handling at the center -- that needed to be done for some time!
Now all of that being said, Gemini 4 Argon does sound pretty cool, and we congratulate both teams at Google -- the team that created Gemini 4 Argon, and the team that refactored Chrome with it!
But all of that being said, while the advancements made for Gemini 4 Argon are important and impressive (and they are!)... refactoring Chrome properly (IMHO) is slightly more important and impressive!
(You know, one man's humble opinion! That is, I am not the guy who is responsible for awarding performance-based bonuses at Google at the end of the year, so take everything I say with the proverbial "grain of salt"! :-) I am probably biased, one way or the other! :-) )
God knows why the server different things to the same people on different interfaces
I never touch 3.8, it's awful. Curious to try this, but I've yet to use any Gemini that is good for agentic work.
Also, unrelated but this made me look at Artificial Analysis to see where it lands on Cost per Task, and it ranks Opus 5.5 at almost double Astra 6. That does not align with my experience.
Google...why. You keep getting so close but constantly fumble at the last moment. I can use Opus 5.5 and Sol/Astra on the $20 Anthropic and OpenAI plans.
You are already playing from behind. You need to GET YOUR MODELS INTO MORE PEOPLES HANDS by any means possible. Continuing to make it difficult to use your products DOES NOT HELP.
This model might be great but the majority aren't going to subscribe for $100/month just to try it.
The fact that multiple companies can breach any benchmark means there is no ambiguity in terms of how to build AI, its just a matter of time.
Certainly Meta and Google are benchmaxxing to please public markets, but it still proves the above
I wonder how much is actual improvements in the model and how much comes from better harnesses. I'd be willing to bet that the combination of better harnesses + the models being trained around agentic workflows (i.e the harnesses) has made it significantly better in many things, but it's hard to find other improvements imo.
A more intelligent model would not default to spaghetti throwing, it would pause to consider the problem fundamentals, the accuracy of assumptions, and ask questions that remove ambiguity. This however does not lend itself to increasing token sales while prices remain inflated from market and brand dominance.