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.
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.
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.
> 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.
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.
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.
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."
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.
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.
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?
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
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.
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.
> 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?
> 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).
Modern tech marketing is a clown car of incompetence.
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.
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.
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.
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.
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.
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?
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
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.
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.
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.
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.
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.
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 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
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.
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".
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.
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.
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
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.
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.
"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.
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.
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!
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 :)
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.
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 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 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.
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
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.
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.
> 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).
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.
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.
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 …
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:
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!
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".
>"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?
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.
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.
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.
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?
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.
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.
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.
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.
They look remarkably good, sure -- but not perfect.
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."
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?
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?
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...
Give me reverse engineering bench and agentic coding.
Leave the visuals to humans which actually have aesthetic sense unlike whatever this mess is
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.
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?
+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).
Modern tech marketing is a clown car of incompetence.
$2/million input and $10/million output today. $4/$20 after December 1st.
Is Elon Musk doing their pricing?
It is price fixing, nothing else
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.
noooooooo
Given FelonyBench, do we also have a FantasyBench?
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.
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.
Do you really think people in google are happy about how long it took to release this model?
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.
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)
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.
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
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.
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.
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.
Is it actually available?
Like a working scroll on a webpage.
God damn this is impossible to use of safari mobile.
The code is bad though.
It feels so complicated and hard to know what we can use vs what hands them our codebases
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.]
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.
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.
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 think these companies have good models but expose them poorly like others have mentioned
i benchmarked 20+ models in real worlds and gemini families are the best (highest customer satisfaction)
> Gemini 4 Argon (Medium)
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.
https://antigravity.google/blog/gemini-4-argon-in-google-ant...
EDIT:
Apparently Ultra subscribers gets Argon everywhere including in Google Antigravity
Pro subscribers only get them in Gemini app and Search
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 crazy if it can work accurately.
Pretty misleading.
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
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.)
Even after using the free model in Google search talking to Anthropic or OpenAI models feels like talking to an ox.
Which one is it?
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.
Edit: noble gases not noble taxes, though the latter would be a funnier naming scheme.
It still thinks Charlie Kirk is alive, though.
Unavailable · Gemini 4 Argon is currently available to Ultra plan subscribers.
"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).
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.
Start preparing Gemma 5 release.
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:
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!
fyi: argon is an interesting choice of name, coming from the Greek argos meaning "lazy or living w/out labor".
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! :-) )
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.
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.
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.