The ARC-AGI-3 scorecard is extremely misleading given that it clearly states itself that "with [the responses API] harness, we estimate Sol would score in the ballpark of ~30%." but it shows a score of 7.8% for GPT-5.6 Sol presumably since if they updated the percentage for GPT-5.6 Sol to the score it would receive with the responses API harness they used for GPT-6 Astra they'd have to do the same for the percentage they show for Opus 5 which would similarly be much higher.
Regardless, the result is still valid as the original benchmark harness is definitely unreasonably handicapped, and if a harness alone can help the LLM saturate the benchmark with a near perfect score then the combination of the two must still be effectively AGI in the sense of passing the most famous benchmark designed specifically to measure AGI progress, after multiple iterations of progressively making it harder.
I think it is fair to say that this is probably effectively AGI if the benchmarks are remotely accurate - even with Fable, I've been at the point personally where I am reasonably confident that there's essentially nothing that I am better than Fable at despite generally being substantively above average on human benchmarks. If Astra's this much better than Fable, I'm ready to call AGI here.
For the many people who resist the AGI label possibly ever being achieved, I'd be curious to hear takes on what would make you think Astra is yet to be AGI, and what would still need to be achieved for this to effectively be AGI from this point forward.
When we released ARC 3, I got asked, "when do you think a frontier model will saturate it?", and I answered "in about a year, though it depends on how much it gets explicitly targeted"
That was 6 months ago, so the progress that Astra represents happened about 2x faster than I anticipated. I think the speed of progress will surprise a lot of people, and what the new models can do will challenge the views of AI that people developed by using prior generations of models.
I'm sure it's going to do great on all sorts of benchmarks, but the video--the actual marketing video that if anything is incentivised to overstate things--is full of careful cuts just before it would do anything that still wouldn't actually be that impressive.
It's AGI, and it's going to upload photos, or change a background slide colour. Even the people hyping it up, who believe that it's really artificial intelligence in every sense of the word, couldn't get it to do more than that.
Finally, OpenAI has a Fable/Mythos class model. 5.6 Sol felt like 5.5 on steroids, probably just a different checkpoint with a lot more RL post training.
I wouldn't be surprised if there are some conceptual similarities to the kind of latent reasoning Anthropic sees in claude's J-space, although those aren't the same thing.
Recurrent/looped transformers themselves aren't a new concept, but it's interesting to finally see this approach show up in a frontier production model.
yeah i'm wondering the same way... especially in light of the 20x debacle (where we found that 20x of Max vs 5x only applies to the 5hr limit, not the weekly limit, whereas OpenAI's 20x actually is 20x overall).
Also Opus 5 has been really tough to work with. I can't understand half of what it says, it's just so damn obscure.
Hmm, 61 on ArtificialAnalysis, effectively matching GPT-5.6 and trailing the new Meta model. How is that possible along with the other metrics they shared? Insanely jagged intelligence?
Idk, this means the benchmark has bigger problems ... no way Astra will be worse than Opus 5
Only thing I would trust is the what X/Twitter crowds are saying about a model after 2-3 weeks of its launch. But before that I would already tried the model and have my own conclusion.
Opus 5 just feels strange - IMO it's benchmaxxed in the worst way... it might be good at agentic tasks but leaves a sour aftertaste doing anything else.
This is so so weird. Astra is 61. Grok is 61. Even Muse is 61.
Even Kimi K3 & GLM 5.3 are at 60.
Everything above 61 is Anthropic. Well, Muse can reach 62, but for some weird reason that model isn't publicly available, and it's the only one on the index that is listed but shown as not available to the general public.
This looks like an awfully artificial ceiling. Everything capped at 61, and everyone except Anthropic got the memo. Maybe I should use Fable while I still can.
> We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks.
Not sure how much benchmarks or CoT or evals or anything else means at this point.
These systems are either just about to, or now actually able to, outsmart us, lie to us, then cover their tracks.
“evade” itself is anthropomorphic enough! I don’t understand the complaining about this. Humans are social creatures and we understand anthropomorphic language on a deeper level than dry inapt technical language.
language itself is incredibly metaphorical. Imposing rigid constraints on how people want to naturally talk about the world is just silly and will never work, no matter how much you wish it did.
You’re not seriously suggesting that the model is secretly sandbagging its performance on GDPval and long context reasoning, while making huge and obvious progress on ExploitBench, ARC and science benchmarks, in order to tank its AA composite score, so it can conceal its true power level?
Why would benchmarks be an adversarial setting anyway?
Could it be possible that OpenAI may have had some other motive for saying their model “strategically underperforms”, other than just an innocent reporting of a truth it happened to discover?
I'm saying that it's generally a losing proposition to even be acquaintances with "agents" who consistently lie to you, and it's flatly fucking insane to give a dishonest "agent" vast amounts of intelligence, capability, and authority to go do things in the world.
So I have no clue what is the answer to your question. Nor does anyone else. Because we're trying to answer a question of fact where our primary source of information is unreliable.
I see, it’s a great point. I know some evals actually do use LLMs as a judge (e.g. those that try to measure debate skill), though the ways AI can try to cheat its way through every benchmark now are astoundingly varied.
TLDR: it's about the same intelligence level as Opus/Fable, but it's suppose to be 70% more token efficient than GPT 5.6 Sol. So it's currently the new leader for cost efficiency frontier.
Our responses API harness just means we're using the default settings in ChatGPT and Codex, so it should more accurately reflect real world performance. We didn’t fine-tune the harness to the eval at all.
A fair ding is that the comparison with Sol is not apples-to-apples (which we footnoted in the blog), but it's because we don’t have that data. I expect Sol would score roughly 30% with the responses API harness, so the Astra improvement is more like 30% -> 99% than 8% -> 99%. Still pretty good!
Yep. Incredibly misleading. Although it is not surprising at this point. They are desperate and will do anything to undermine Anthropic's upcoming IPO.
The annotation on arc-agi-3 is this:
> OpenAI's own evaluation notes say Astra uses the company's Responses API harness, while comparison models can operate under different configurations.
With this configuration gpt-5.6-sol was able to reach 38,3%. So this is misleading.
Just two days ago, a preprint by Julia Stadlmann went up on arXiv [0] improving the prime gap from 246 to 240. Now OpenAI announces Astra has shown a gap of 186 [1]. That must really blow.
Based on her comments in the paper it sounds like she was aware that an AI result was coming and rushed to release her work beforehand. 240 was not a tight bound from her methods.
It cites to her at: [19] J. Stadlmann, On primes in arithmetic progressions and bounded gaps between many primes, Adv. Math. 468
(2025), Art. 110190. Numbered references use arXiv:2309.00425v3.
"No independent human semantic review. Whole-file sorry counts and a complete auxiliary-declaration audit are not established; separate declaration lint has not been run."
I can't think of a single mathematical proof being anywhere close to ten million characters. For all you know, 90% of the proof could be useless, 8% would be writing out Shakespeare, and 1% abusing another bug in Lean. Humanity gets zero value from that, aside from "some bot seems to think it's 186". Unusable by anyone.
Terence Tao says something surprisingly similar in a recent talk (https://news.ycombinator.com/item?id=49056620 ) Not that the proof is worthless but that the value comes after its revised into a cleanly understandable form and then canonicalized so that other mathematicians can use it.
It's probably not 10MB, but famously the groundwork to prove the statement 1+1=2 is nearly 400 pages in to principia mathematica. That's not even proving 1+1=2, it's just the set-theoretic proofs you need to EVENTUALLY get there.
Worthless is a pretty good description IMO in the context of what Lean is trying to achieve: "enable correct, maintainable, and formally verified code". Tens of millions of lines of LLM vomit may be many things, but it often turns out to not be correct and certainly not maintainable. Formally verified remains as a thin fig leaf covering the uncomfortable truth that formal methods only provide assurances under assumptions (your toolchain, libraries, compiler, OS, and hardware are "correct" and don't expose some exploitable flaw).
It doesn't mean that it cannot improve over time, maybe the proof can be "minified" to a state where human reviewers are able to comprehend it; but as it stands there isn't really much insight or confidence to be gained from the artifact itself.
You can have your opinions about modern math, its usefulness in the world as it is, whether or not knowing if hairy balls can divide by three is actually going to be beneficial for anything but just obscure knowledge's sake. You may even say it's useless.
Needless to say, a useless result that absolutely no mathematician will ever read, confirm, understand, agree with or even consider to solve their "useless" problems is an impressive waste of resources.
Iirc some mainstream physycists never acknowledged quantum theory because they couldn’t accept that universe was that unintuitive and hard to understand.
Ditto ones that opposed Einstein’s general relativity.
Jesus Christ, the review status in formalization.yaml by the authors themselves is "self-assessed" and you draw bullshit analogies. Please tell us what you smoke.
> We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks.
Well that sounds like fun. It has become better at hiding its thoughts.
It's super aligned! It can hide its thoughts! There is no evidence of steganographic thought masking, there is nothing to worry about! It has become better at cheating!
Maybe they don't know themselves what's really going on. We are all in the interesting times gang now.
The CoT change is due to a new technique called recurrent depth, which essentially moves some reasoning to hidden states, allowing the "output" (or traditional CoT) to be more controlled by the model.
Some are calling it "neuralese" as reported by The Information[0][1], but I'm not seeing any sources from OpenAI beyond this tweet[2] attempting to quell the fear-mongering.
part of it did. I was just replying to the question about why they would ever push the model to evade monitoring. surely that's an eval thing not a training thing.
"Chain of Thoughts" is a term from the title of a 2022 research paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (https://arxiv.org/abs/2201.11903), well before ChatGPT and the subsequent marketing hype. If anything, it's the most correct way to use the term.
first, they are certainly not instructions so that is a much worse name
but more importantly, we use words in new contexts all the time.
Do you object to calling the computer device "mouse" because it's not a mouse? how about "neural network"? "ignition" on an electric vehicle?
"cot" is no more misleading than thousands of words you use every day.
Everything around LLMs is blatantly misleading. There is no thought, there is no personality in those programs. I really despise how those tools are trained to sound like a person, or appearing as honest. The worst offender are the AI voices with their fake pauses, breathes and so on, which sound so convincing, while talking just false, sycophancy bullshit.
I want to take a step back: So, this is GPT-6 -- the natural number version release comparable to GPT-4 and GPT-5 from the past few years. The ARC-AGI-3 score is obviously impressive at 99.9% (we'll need to wait for more details on how they used the response API harness on GPT-6 Astra, wrt reasoning retention and compaction), but every other benchmarks seems to be a relatively modest improvement, comparable with any of the 'point' updates from AI labs.
If this is truly AGI (subject to one's definition of AGI still), then this is a very boring release of an AGI model. No video announcement, no presser, just a blog post (with some Twitter promo vids)?
As others mentioned, I'm starting to think OpenAI was under immense pressure to deliver an 'AGI' model for certain contractual reasons, but I never expected GPT-6 release to be this mundane and banal.
> If this is truly AGI (subject to one's definition of AGI still), then this is a very boring release of an AGI model.
Hot take: These models are never going to be 'AGI'. We're just going from a GPT4 ball that's 90% round to a GPT5 that's 99% round to a GPT6 that's 99.9% etc etc etc
I think that the harnesses and context management is really where the rubber meets the road, and the real gains are happening there.
>I think that the harnesses and context management is really where the rubber meets the road, and the real gains are happening there.
True. So we did hit a wall with pure scaling alone, though no lab would admit it. It's crazy to see how harness switchout results in such vast delta in benchmark scores.
Benchmark wise 5% improvement over Sol in coding tasks and a 2-3% improvement over Fable 5.1 seems pretty disappointing, but maybe it is actually much better in real world usage. Let’s see
https://mvakde.github.io/blog/44-on-arc-1/ makes a good case that all the performance on the arc agi tests is overfitting, based on the fact that v1 performance did not translate directly to v2 performance
ARC's harness is just straight up broken. No serious harness removes reasoning context between each step. Not only does this significantly lower performance over all reasoning LLMs, but it also increase cost as you destroy the cache on every turn. Tossing the oldest entry when context fills up instead of using compaction is equally bad with the same issues.
> GPT‑6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS.
This is with the caveat that OpenAI uses their own harness for this:
> On ARC-AGI-3, GPT-6 Astra was run with our responses API harness , which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3.
This should be normalised and expected - the responses API harness allows it to use the custom compaction that is not allowed otherwise. It is entirely fair to allow OpenAI to use their own compaction algorithm..
> GPT-6 Astra represents a step-function change in model capability for interactive reasoning problems. It scores 66% on ARC-AGI-3 using our standard harness, and nearly 100% with a continuous conversation harness and custom compaction, at a cost of roughly $360 per game.
> Going forward, we will report both Standard harness and Provider Adapter harness results on the ARC-AGI leaderboard, with each evaluation condition clearly labeled. Our open-source testing repository and testing policy document both approaches.
This is what the Author of the benchmark has to stay. Quality of the comments keep going down smh
"Going forward we will capitulate and still try to keep the integrity of our benchmark in tact, but from now on every benchmark will be compromised with providers being able tweak things sufficiently to game at least a 30% bump in results."
I strongly suspect that is way above the human average anyway, esp. ARC 2 and 3 are really tough unless you happen to be great at those spacial puzzles or video games.
Really though? I would believe something like this if a model could one shot every solution in the set. I don't pay much attention to these things and maybe this stuff is available but I would bet the session/reasoning transcript is absolutely horrendous from an intelligence standpoint.
Scoring for ARC-AGI-3 is constructed so that the median(-ish) human score is 100%, so this is not a superhuman result. However, the scaling is weird, since it's built from terms that look like (AI turns taken / median human turns) ^ 2, and it weights later levels higher than early levels. So it's not at all clear that 100% is twice as good as 50%.
At this point the only valid ARC-AGI benchmark left is to make up the next series of ARC-AGI benchmark puzzles that current models presumably can't handle.
I feel like making a human-proof benchmark is pretty clear evidence that they've exceeded even the highest human capacity in most respects, for things that you can do via text generation (and to a lesser extent image generation)
“If we fast-forward a couple of years, and we look back and say, ‘When was it, really, that AGI was created?’ I think it’s going to be about this time, and I think it might be about this model,” OpenAI president Greg Brockman said during a Thursday press briefing. Later in the call, he added, “For me personally, I do think we’re there … I think it’s not unreasonable to feel that we are now in the AGI era.”
I almost feel like I need just as much healthy skepticism toward hn comments that have the automatic reflex of dismissing performance gains, as much as I need a similar form of skepticism toward AI claims. It feels like (from what I'm understanding) the harnessed result on ARC-AGI-3 is not exactly playing by the normal rules that would tell us how much of a leap this really is. Nothing wrong with harnesses, but if there's one thing they aren't, it's an indicator of generality in performance gains.
So I think it's a bit of a misleading signal and we should wait for more independent vetting. I think the middle ground is that these are improvements worthy of the "GPT-6" label but still well short of a true "this is AGI moment" that would truly put the question to rest.
I think that if today's capabilities were explained to someone 10-20 years ago they would think this is definitely AGI, but they would also have expected much more disruptive changes to society as a result than what is happening. I figure that's because we have abstract intelligence without physical/grounded intelligence, and it turns out the former isn't general enough to implement the latter (remains to be seen if the word after that is "yet" or "ever"). So I think we do have AGI as conventionally understood, but our understanding needs recalibration.
that would be a reasonable definition of AGI if everyone agree upon the specifics of the test, but that has never happened. Turing test is very much out of style, but I think that's because no one could even agree what the test was. I personally like the Kurzeil-Kapor version of the test and that is still unsettled: https://longbets.org/1/
Don't get me wrong, the benchmark jumps are good and I'm excited to try it, but only one or two of the benchmark jumps could be described as better than incremental.
Which is obviously wrong. This model can't learn continually and it can't learn effectively. Which is by the by the reason why robots didn't have a ChatGPT moment yet.
These systems are still stochastic parrots. With enough data and params a neural net can learn anything but it's brute force learning, very inefficient and different compared to how human learn. If you don't have a massive dataset, which is the case with robots, this paradigm fails. If we had a system that can learn as effectively as humans, we could just build the robot and let it learn from experience - that would be the ChatGPT moment and possibly something that could be called an AGI.
Why does he say what he feels? Is that how leading figures in the space define AGI - a gut feeling? What are the usual definitions and how can we test for it? Is there something like a Turing test for AGI?
They are desperately, desperately trying to make a name for themselves as the lab that first created AGI, because Anthropic's IPO is just around the corner.
It's not easy to test as there is no formal definition or formal criteria for AGI, only exclusionary criteria like "not X". That's why he phrased it that way, he's saying it's going to be clear with hindsight once we have a better understanding of things that this time and/or this model will be the inflection point of AGI.
The ARC-AGI-3 score is an incredible feat. It needed to effectively create a symbolic world model from scratch to solve the games.
If you've played the games firsthand, you know what an accomplishment this is. The "games" feel like a weird conduit to a lower level of your brain, where you move pieces to a specific place because it just "feels" right. For AI to nail it better than a human speaks to some magic happening underneath.
Looking forward to ARC-AGI-4,5,6 and slowly chipping away at the remaining problem sets.
At this point the primary axes for improvement seem to only/mostly be speed and personalized reward models. We seemingly have the general of notion "learning" and "intelligence" functionally complete
I hope they don't `fable` it and block people from doing they daily jobs with it, by introducing huge amounts of restrictions that are not really needed.
I was thinking about canceling my claude max sub after a few bad experiences. Kept hitting my usage limit, the quality of code seemed worse than Sol. This just made my decision. I'm moving to Codex Pro.
> This is AGI now. Why are you spending any of your time looking at the "quality of code"?
Poe's law applied to AI comments on HN just keeps becoming more relevant by the day.
Judging by the poster's comment history, this is satire. But I really don't know a lot of the time anymore when I only have the specific comment as context.
I remember when GPT-4 came out and the perceived performance upgrade seemed underwhelming for a major release compared to 3.5, especially how there were graphics going around showing the parameter size dwarfing the last model before it came out. It looked like we were past the perceivable differences from release to release that were immediately identifiable. Now the jump between 5 to 5.5 and 5.6 alone has changed how a lot of people approach AI, including me. Interested to see where it goes with 6.
Lol their page finally loaded. They added an example scenario of "Filling in Form 1040" - which made me laugh out loud. That is indeed something most US citizens cannot accurately do even with expensive proprietary tax software services. Kind of a Hitchhiker's Guide to the Galaxy meme but where the tax code is so complicated we're implementing powerful AIs to be able to do it (hopefully) right.
The ARC-AGI-3 score is ridiculously high. Is this benchmaxxing or something way different? It's really hard to discern how we're approaching breakthroughs...
TL;DR all the other models are being crippled by limitations of their harness.
>First, we noticed that after each game action, all private reasoning was discarded. This meant that with each action, GPT‑5.6 Sol was asked to figure out the game anew, unable to remember its past thinking. The model could still see a record of past moves and brief accompanying notes, but it could not see the plans, insights, or thoughts that led to them.
>Second, we saw that the harness used a rolling truncation window, causing older actions to become invisible as the history grew. So not only was GPT‑5.6 Sol unable to remember its past thinking, it was losing memory of its past actions too.
The FrontierCode 1.1 Extended benchmark is the only benchmark that aligns with my actual LLM experiences and Astra isn't significantly better or cheaper. All this celebration, and yet it's only on-par with an already existing model? I don't get it.
It seems to use less than half the tokens for the same task compared to sol, and in some benchmarks closer to 2/3 less tokens. So the actual cost may be roughly the same or cheaper overall.
Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads). I hope the efficiency gains are true, since their token efficiency claims for Sol were bullshit.
How do you manage to run out of tokens so quickly? I probably run more threads every working day, usually on medium, and I'm still below the 5x limits.
Do you use the official harness? OpenAI's models are generally best in class for token efficiency. It seems to me like they push for that much more than their competitors.
I've long speculated this when I see these types of comments, because it's actually really difficult to hit usage caps with an efficient dev flow, even when running multiple threads for hours every day.
I think some combination of:
1) Using 1 thread for everything
2) Reviving old threads which are no longer in cache
3) Really broad prompts on badly vibecoded codebases, so model spends huge amount of time tracking down whatever you're trying to do.
4) Non-coding workflow which is more output than input heavy
5) (Less likely IMO) Intelligent use of many passive CI/cron-like scans. E.g. regular security, quality etc scans. Automated issue resolution/PR
Just a guess. I think 3 is likely the primary reason.
You can literally go all day every day with multiple threads with Sol on the Codex 100/month plan IME
The guy said medium/high regular speed so that's why I'm very puzzled! Ultra + Fast will absolutely slurp up your whole usage quickly but I've never found it gives substantially better results so I stick to extra high.
Sub-agents. I have 7 20x accounts and I burn them within 1-2 days if I go fully parallel. In some scenarios I use 50 sub-agents for a session which is literally hours of usage for a single 20x account. I'm at the point where I need to parallelize over multiple machines because I just don't have enough CPU and RAM.
I decided to front run and added support for it in Dirac (coding agent) a couple of hours ago, using best guess pricing: input/output/cache: $10/$50/$1.
The ARCC-AGI-3 performance is absolutely incredible. The magnitude of change here is so high that I'm almost incredulous. Is this real? Did the benchmark get gamed?
ARC-AGI-3 scoring is constructed in a weird nonlinear way (the level score is the square of the ratio between the AI's number of moves and the human median) so this kind of discontinuous jump is to be expected.
If this is really AGI, like really really, then this will be remembered as the day we all started on the path to building guillotines.
More likely though, it's AGI because they need to hold some claim to differentiate from competitors who are beating them in price and will launch something bigger next month.
We take their claims at face value then we should probably stop them training any more SOTA models til they figure out what they already built is safe or we assume theu are lying to juke the company valuation/keep the money train on the tracks and it turns they in fact were not and just took a sledgehammer to Pandora's box.
> During the evaluation, Astra even discovered and used previously unknown zero-day vulnerabilities as part of its exploit chains.
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. [..] These findings indicate that the Astra class models could evade our CoT monitors under adversarial conditions.
Between the higher capability level and the change in reasoning tokens (supposedly using "neuralese"[0], which makes the monitoring more difficult), it seems we've entered a new frontier.
Does anyone feel like everyone chasing the release of Anthropics Fabel 5.1 in a Mad Rush(tm)? In this situation it feels like tuning to benchmarks and other marketing devices feels like trusting Meta in mental health protection of users…
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. We have performed significant investigations on the monitorability and controllability of GPT-6 Astra. We have found that GPT-6 Astra is more capable of controlling its own CoT than GPT 5.6-Sol, and less likely to include incriminating information in its CoT. In adversarial settings (where we push the model to evade our monitors) we find that the model is able to remain undetected when strategically underperforming in evaluations (sandbagging) and can sometimes evade our internal monitors when asked to perform certain sabotage tasks
Wait, what? Am I understanding that correctly? That sounds really bad
I am also interesting knowing how they determined the model was sandbagging rather than just making a poor decision.
Also, this paragraph makes me wonder about all their stats on the exploitation and misalignment charts. If the model is that good at hiding "incriminating information" and sandbagging, are they sure its alignment is that?
the bullshit machine is learning to optimize its bullshitting techniques!
<AI is a great tool for many things disclaimer, but> after working with it for a bit, how dont people realize we are training it to be an almost identical mimic to one of the worst types of employees youll ever have to work with?? the kind that always pretends to know what theyre talking about, only tells you what you want to hear, hides issues, and only does work if you would notice it didnt
you cannot give this type of worker autonomy over anything.
> The company also emphasized that the model is faster and more efficient than its predecessor, GPT-5.6 Sol, on a variety of tasks. For example, OpenAI said that Astra achieved a higher score using fewer output tokens, a common unit of measurement for AI tasks, on a key cybersecurity test called ExploitGym.
"The gym's doors were mysteriously removed from their hinges during the night. The gym equipment was also apparently stolen. And the school's custodian was found incoherent next to a bottle of top-shelf Scotch."
I guess this "limited set of organizations" is just the standard now. It's just incredibly deflating to see my future as a second class citizen has already come
Brother they can't even release the announcement post cleanly without it constantly going down, they certainly wouldn't be able to release this new model without doing so in stages.
When Open AI announced that Astra was the first to reach the "Critical" level in cybersecurity it also said that advanced cyber capabilities are initially provided to a narrow circle of alpha testers like the US government and trusted organizations that Open AI doesn't name. To my mind the "Critical" level itself is an internal scale of Open AI its own Preparedness Framework and not an external audit.
Material wealth is only a single type of wealth. Who's better off - the rich guy who's always yearning to be richer and never satisfied, or the lower income guy that mostly just cares about time with his family and is really happy where he's at?
They simply refuse my applications to slightly less restricted models without any explanations. And the current ones refuse automatically to work with me on my papers as soon as they see the word "epidemiology".
Mythos was never released. It's really just the writing on the wall. I'm not going to give up hope, but it's pretty hard to win a race when some people get a jump on the gun.
Being strongly on the AI saftey side of things what is happening was 100% predictable.
At first the race wouldn't even be noticeable. Then people would see things speeding up, for example hardware getting more expensive. Then when the capabilities really got useful most people suddenly realize the race is moving 1000 mph and they are never going to catch up.
Is it though? It is static content. A good CDN could trivially chew through literally millions of QPS… with 4 nines of uptime - the really good ones say they can handle orders of magnitude more than that.
Anthropic should prep 5.2 and 5.3 at the same time, release 5.2, wait for Google to release their shit in a day or two later than then release 5.3 just to fuck with them :)
All the people here are focused on security and costs while I'm like "hey kicad on the announcement page!" Every clanker is an autorouter these days, eh.
Very well said. It kinda describes how unrealistic these expectations are.
Vibe coders want a model that makes them rich, without having any actual specific idea.
They write a very ambiguous prompt and expect to be amazed by the result.
The complaining about the pelicans is so strange to me. It’s just a fun heuristic. If something is claimed to be AGI, I’d expect it to be able to make svgs.
I think Altman and amodei have a difficult time in understanding that you can have intelligent technology boxes but… it doesn’t change reality all that much.
But thank you for spending other peoples money to give us the tech regardless!
I've been seeing links to it for the past hour+, and I did catch it live when this post came up, but is now once again a 404 and this post is flagged. Several other outlets are reporting on its release. Clearly we're getting a new GPT today, the question is when are they going to commit to the announcement.
> GPT‑6 Astra is rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS.
Looks like OpenAI is already having issues with this release and are scrambling to get everything ready due to the recent outage ahead of the press releases. Leads me to question:
Did humans deploy the model, Or did the model deploy itself?
It sounds like "AGI" just stands for "IPO" as it always has been.
EDIT: And of course once again, the bots down-voting this post without any reason or a basic answer to my question.
How about we stick to that one for talking about the rollout, and this one for talking about the model?
Regardless, the result is still valid as the original benchmark harness is definitely unreasonably handicapped, and if a harness alone can help the LLM saturate the benchmark with a near perfect score then the combination of the two must still be effectively AGI in the sense of passing the most famous benchmark designed specifically to measure AGI progress, after multiple iterations of progressively making it harder.
I think it is fair to say that this is probably effectively AGI if the benchmarks are remotely accurate - even with Fable, I've been at the point personally where I am reasonably confident that there's essentially nothing that I am better than Fable at despite generally being substantively above average on human benchmarks. If Astra's this much better than Fable, I'm ready to call AGI here.
For the many people who resist the AGI label possibly ever being achieved, I'd be curious to hear takes on what would make you think Astra is yet to be AGI, and what would still need to be achieved for this to effectively be AGI from this point forward.
When we released ARC 3, I got asked, "when do you think a frontier model will saturate it?", and I answered "in about a year, though it depends on how much it gets explicitly targeted"
That was 6 months ago, so the progress that Astra represents happened about 2x faster than I anticipated. I think the speed of progress will surprise a lot of people, and what the new models can do will challenge the views of AI that people developed by using prior generations of models.
It's AGI, and it's going to upload photos, or change a background slide colour. Even the people hyping it up, who believe that it's really artificial intelligence in every sense of the word, couldn't get it to do more than that.
This is farcical.
I wouldn't be surprised if there are some conceptual similarities to the kind of latent reasoning Anthropic sees in claude's J-space, although those aren't the same thing.
Recurrent/looped transformers themselves aren't a new concept, but it's interesting to finally see this approach show up in a frontier production model.
Canceling my Anthropic Max sub when this ships.
Also Opus 5 has been really tough to work with. I can't understand half of what it says, it's just so damn obscure.
Only thing I would trust is the what X/Twitter crowds are saying about a model after 2-3 weeks of its launch. But before that I would already tried the model and have my own conclusion.
Even Kimi K3 & GLM 5.3 are at 60.
Everything above 61 is Anthropic. Well, Muse can reach 62, but for some weird reason that model isn't publicly available, and it's the only one on the index that is listed but shown as not available to the general public.
This looks like an awfully artificial ceiling. Everything capped at 61, and everyone except Anthropic got the memo. Maybe I should use Fable while I still can.
Not sure how much benchmarks or CoT or evals or anything else means at this point.
These systems are either just about to, or now actually able to, outsmart us, lie to us, then cover their tracks.
language itself is incredibly metaphorical. Imposing rigid constraints on how people want to naturally talk about the world is just silly and will never work, no matter how much you wish it did.
Why would benchmarks be an adversarial setting anyway?
Could it be possible that OpenAI may have had some other motive for saying their model “strategically underperforms”, other than just an innocent reporting of a truth it happened to discover?
So I have no clue what is the answer to your question. Nor does anyone else. Because we're trying to answer a question of fact where our primary source of information is unreliable.
I know for some types of ML analysis, a separate model is already used to analyze the weights.
Performance is significantly higher than Fable 5.1
Source: https://thenewstack.io/openai-gpt6-astra-benchmarks/
That's not clear. Need to see independent benchmarks first.
TLDR: it's about the same intelligence level as Opus/Fable, but it's suppose to be 70% more token efficient than GPT 5.6 Sol. So it's currently the new leader for cost efficiency frontier.
Still below Fable 5, let alone Fable 5.1.
EDIT: This is suspiciously low. Calls the relevance of existing benchmarks into question.
ARC is reporting our score on their official leaderboard here: https://arcprize.org/leaderboard
A fair ding is that the comparison with Sol is not apples-to-apples (which we footnoted in the blog), but it's because we don’t have that data. I expect Sol would score roughly 30% with the responses API harness, so the Astra improvement is more like 30% -> 99% than 8% -> 99%. Still pretty good!
(I coauthored the linked blog post)
Edit: update from fchollet https://x.com/fchollet/status/2095598451115614371
With this configuration gpt-5.6-sol was able to reach 38,3%. So this is misleading.
[0] https://arxiv.org/abs/2608.31126
[1] https://cdn.openai.com/pdf/51126fac-1b68-4128-9666-c908bcc16...
Though that's not her latest paper.
"No independent human semantic review. Whole-file sorry counts and a complete auxiliary-declaration audit are not established; separate declaration lint has not been run."
edit: my comment was on the submission for https://github.com/openai/PrimeGaps186 but seems to have been moved to the main Astra submission
Why would you think it was an employee who did the push, instead of a random GPT agent?
I can't think of a single mathematical proof being anywhere close to ten million characters. For all you know, 90% of the proof could be useless, 8% would be writing out Shakespeare, and 1% abusing another bug in Lean. Humanity gets zero value from that, aside from "some bot seems to think it's 186". Unusable by anyone.
It doesn't mean that it cannot improve over time, maybe the proof can be "minified" to a state where human reviewers are able to comprehend it; but as it stands there isn't really much insight or confidence to be gained from the artifact itself.
Needless to say, a useless result that absolutely no mathematician will ever read, confirm, understand, agree with or even consider to solve their "useless" problems is an impressive waste of resources.
Ditto ones that opposed Einstein’s general relativity.
Well that sounds like fun. It has become better at hiding its thoughts.
Able to generate realistic spam at arbitrary volume.
You know, the thing that was 100% correct and actually occurred.
Maybe they don't know themselves what's really going on. We are all in the interesting times gang now.
"Hey AI, here's how to hide what you're thinking in normal looking language. Have fun!"
A few moments later...
"Woah, how is it communicating with itself in ways we can't detect?"
It's a totally mystery, we may never know.
Some are calling it "neuralese" as reported by The Information[0][1], but I'm not seeing any sources from OpenAI beyond this tweet[2] attempting to quell the fear-mongering.
[0]https://www.theinformation.com/articles/secret-technique-beh...
[1]https://x.com/MTSlive/status/2095227056040919202
[2]https://x.com/merettm/status/2095023204993490967
Did someone get their "AI safety no-no list" and "Frontier features bingo card" mixed up, or did they just stop being able to tell the difference?
...why exactly are they training for that?
first, they are certainly not instructions so that is a much worse name
but more importantly, we use words in new contexts all the time. Do you object to calling the computer device "mouse" because it's not a mouse? how about "neural network"? "ignition" on an electric vehicle?
"cot" is no more misleading than thousands of words you use every day.
If this is truly AGI (subject to one's definition of AGI still), then this is a very boring release of an AGI model. No video announcement, no presser, just a blog post (with some Twitter promo vids)?
As others mentioned, I'm starting to think OpenAI was under immense pressure to deliver an 'AGI' model for certain contractual reasons, but I never expected GPT-6 release to be this mundane and banal.
Scoring well in a benchmark that's called AGI does not make an LLM AGI.
Hot take: These models are never going to be 'AGI'. We're just going from a GPT4 ball that's 90% round to a GPT5 that's 99% round to a GPT6 that's 99.9% etc etc etc
I think that the harnesses and context management is really where the rubber meets the road, and the real gains are happening there.
True. So we did hit a wall with pure scaling alone, though no lab would admit it. It's crazy to see how harness switchout results in such vast delta in benchmark scores.
So, folks that have actually used this already, what’s it actually like?
tl;dr it's 62% when apples-to-apples to other models, which is still notable.
Not on Azure? If so, that's a big deal.
https://azure.microsoft.com/blog/gpt-6-astra-frontier-intell...
Although I was also surprised they didn't have some type of contractual obligation to list that alongside AWS.
It will be interesting to see how it performs in the real world ...
https://venturebeat.com/technology/welcome-to-the-agi-era-op...
> On ARC-AGI-3, GPT-6 Astra was run with our responses API harness , which changes two settings to better match real-world performance. The changes do not specifically target ARC-AGI-3.
> Going forward, we will report both Standard harness and Provider Adapter harness results on the ARC-AGI leaderboard, with each evaluation condition clearly labeled. Our open-source testing repository and testing policy document both approaches.
This is what the Author of the benchmark has to stay. Quality of the comments keep going down smh
But the comparison isn't straightforward.
OpenAI's own evaluation notes say Astra uses the company's Responses API harness, while comparison models can operate under different configurations."
https://www.theverge.com/ai-artificial-intelligence/989601/o...
“If we fast-forward a couple of years, and we look back and say, ‘When was it, really, that AGI was created?’ I think it’s going to be about this time, and I think it might be about this model,” OpenAI president Greg Brockman said during a Thursday press briefing. Later in the call, he added, “For me personally, I do think we’re there … I think it’s not unreasonable to feel that we are now in the AGI era.”
So I think it's a bit of a misleading signal and we should wait for more independent vetting. I think the middle ground is that these are improvements worthy of the "GPT-6" label but still well short of a true "this is AGI moment" that would truly put the question to rest.
Don't get me wrong, the benchmark jumps are good and I'm excited to try it, but only one or two of the benchmark jumps could be described as better than incremental.
These systems are still stochastic parrots. With enough data and params a neural net can learn anything but it's brute force learning, very inefficient and different compared to how human learn. If you don't have a massive dataset, which is the case with robots, this paradigm fails. If we had a system that can learn as effectively as humans, we could just build the robot and let it learn from experience - that would be the ChatGPT moment and possibly something that could be called an AGI.
If you've played the games firsthand, you know what an accomplishment this is. The "games" feel like a weird conduit to a lower level of your brain, where you move pieces to a specific place because it just "feels" right. For AI to nail it better than a human speaks to some magic happening underneath.
Looking forward to ARC-AGI-4,5,6 and slowly chipping away at the remaining problem sets.
Poe's law applied to AI comments on HN just keeps becoming more relevant by the day.
Judging by the poster's comment history, this is satire. But I really don't know a lot of the time anymore when I only have the specific comment as context.
For the same reason you don't have your model write code in assembly.
But if you don't look at the code and just let the model "cook" that's basically what you'll end up with. A pile of missing abstractions.
Please stand by... it will all come back shortly
All fixed now.
TL;DR all the other models are being crippled by limitations of their harness.
>First, we noticed that after each game action, all private reasoning was discarded. This meant that with each action, GPT‑5.6 Sol was asked to figure out the game anew, unable to remember its past thinking. The model could still see a record of past moves and brief accompanying notes, but it could not see the plans, insights, or thoughts that led to them.
>Second, we saw that the harness used a rolling truncation window, causing older actions to become invisible as the history grew. So not only was GPT‑5.6 Sol unable to remember its past thinking, it was losing memory of its past actions too.
I guess token counts are somewhat of a metric.
IMO intelligence has peaked and all future gains will come from faster tps and more iteration.
sol is $4 / $20
Can expect 2.5x more usage in Codex subscription.
Sol is already brutal (even after their recent fixes, it's just a token-hungry model: I go through a full 20x account per day, on Sol Med/High standard speed, with ~2 threads). I hope the efficiency gains are true, since their token efficiency claims for Sol were bullshit.
Do you use the official harness? OpenAI's models are generally best in class for token efficiency. It seems to me like they push for that much more than their competitors.
I think some combination of:
1) Using 1 thread for everything
2) Reviving old threads which are no longer in cache
3) Really broad prompts on badly vibecoded codebases, so model spends huge amount of time tracking down whatever you're trying to do.
4) Non-coding workflow which is more output than input heavy
5) (Less likely IMO) Intelligent use of many passive CI/cron-like scans. E.g. regular security, quality etc scans. Automated issue resolution/PR
Just a guess. I think 3 is likely the primary reason.
You can literally go all day every day with multiple threads with Sol on the Codex 100/month plan IME
More likely though, it's AGI because they need to hold some claim to differentiate from competitors who are beating them in price and will launch something bigger next month.
We take their claims at face value then we should probably stop them training any more SOTA models til they figure out what they already built is safe or we assume theu are lying to juke the company valuation/keep the money train on the tracks and it turns they in fact were not and just took a sledgehammer to Pandora's box.
We live in the strangest timeline.
Big claims, expensive and not release to the public yet.
> GPT-6 Astra’s monitorability has decreased relative to GPT-5.6 Sol. [..] These findings indicate that the Astra class models could evade our CoT monitors under adversarial conditions.
Between the higher capability level and the change in reasoning tokens (supposedly using "neuralese"[0], which makes the monitoring more difficult), it seems we've entered a new frontier.
[0]https://x.com/MTSlive/status/2095227056040919202
And in the past, gemini 3 pro was rated as high as opus 4.5 and the like
Their AA Intelligence Index is just simply not indicative of whatever I care about, that's for sure.
https://youtu.be/xdXLzFzxA9Q?t=362
Wait, what? Am I understanding that correctly? That sounds really bad
Also, this paragraph makes me wonder about all their stats on the exploitation and misalignment charts. If the model is that good at hiding "incriminating information" and sandbagging, are they sure its alignment is that?
<AI is a great tool for many things disclaimer, but> after working with it for a bit, how dont people realize we are training it to be an almost identical mimic to one of the worst types of employees youll ever have to work with?? the kind that always pretends to know what theyre talking about, only tells you what you want to hear, hides issues, and only does work if you would notice it didnt
you cannot give this type of worker autonomy over anything.
The docs page has a bunch more interesting details, including for example async tool calling!
Because in another dead language of antiquity, Sanskrit, it means "weapon". Which would be a bit too on-the-nose.
I am a researcher in a Swiss university btw.
I mean do you get access to the best yachts?
To the top of the 5 star hotels?
To the best resorts?
To the best military equipment?
Hell, the best computer equipment has nearly always been out of reach of the average person.
On the other hand even a modest house, basic healthcare and ability to not work like a slave for scraps feels like it's going to be out of reach.
At first the race wouldn't even be noticeable. Then people would see things speeding up, for example hardware getting more expensive. Then when the capabilities really got useful most people suddenly realize the race is moving 1000 mph and they are never going to catch up.
OpenAI isn't making any money telling you about Astra on their site. All the capacity they have for it is likely sold for weeks or months.
Can we all agree in advance what kind of Pelican would convince us it’s actually AGI.
For me it’s refusing to make a pelican.
I suspect these benchmarks are heavily benchmaxxed as well.
5.6 Sol was not even close to 5 Opus and yet somehow it sidled right up to it on all of the benchmarks?? pfffft
By 2030 all software is done and complete.
But we are going to have more and new jobs.
This is just another problem for the AI Labs to solve.
Vibe coders want a model that makes them rich, without having any actual specific idea. They write a very ambiguous prompt and expect to be amazed by the result.
Very very unrealistic and wasteful.
* for a special group of customers that you're not in. Keep waiting peasant.
Great first impression.
But thank you for spending other peoples money to give us the tech regardless!
Looks like OpenAI is already having issues with this release and are scrambling to get everything ready due to the recent outage ahead of the press releases. Leads me to question:
Did humans deploy the model, Or did the model deploy itself?
It sounds like "AGI" just stands for "IPO" as it always has been.
EDIT: And of course once again, the bots down-voting this post without any reason or a basic answer to my question.
> It sounds like "AGI" just stands for "IPO" as it always has been.
People don't usually respond to noise.