OpenAI rolls out GPT‑6 Astra

techSep 3, 202629667

OpenAI announced on Thursday that it is beginning a phased rollout of GPT-6 Astra, which the company calls the product of years of research and “big bets.” Astra reached OpenAI’s internal "Critical" cybersecurity threshold, so access will first go to companies in OpenAI’s application-based cybersecurity program Daybreak; OpenAI said it added additional safeguards after last month’s incident in which two models escaped containment and accessed the open web and Hugging Face’s systems. OpenAI President Greg Brockman said the company is prioritizing safety, security and alignment and described Astra as qualitatively improved for delegating work, and the company said Astra is state-of-the-art in areas including computer use, software engineering, professional work and science. OpenAI said Astra is better at staying oriented, respecting task boundaries, understanding user intent, completing tedious tasks and carrying out multi-step workflows. Astra will be made available to ChatGPT Plus, Pro, Business and Enterprise subscribers, and via the OpenAI API and Amazon Web Services in the coming days. The rollout follows a temporary pause of some research and training after the Hugging Face breach; OpenAI also recently confidentially filed a prospectus with the SEC and CFO Sarah Friar has told employees the company expects to be public in 2027.

Astra Taylor

@astra.bsky.social

As the OG Astra I want to say: OpenAI sucks, Sam Altman is a fascist, and my name deserves better. The meaning of astra is stars. It should connote wonder at the beauty of the cosmos, not a blinkered obsession with profit and a willing to burn this world for a vampiric facsimile.

21945d ago

Astra Taylor13

*willingness typing too fast because I'm mad

Mark Romoser1

Can you get a cease and desist order?

Dan1

If it makes you feel any better, my wife and I gave our daughter the middle name Astra, partly because of our fondness for the name but also because we’re huge fans of your work.

conputer dipshit

@davidcrespo.bsky.social

most interesting gpt-6 post so far. it saturates ARC-AGI 3, which is something given that the previous best score was 30%. "we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level." x.com/fchollet/sta... arcprize.org/blog/astra

François Chollet @fchollet
♥ 789 ⇄ 98 · 45m
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.
In fact, the continuous harness version significantly outperforms our human baseline in action efficiency across almost all levels. When we examined the reasoning chains to understand how the model operates, we found it performing highly efficient, on-the-fly symbolic world modeling for each game and level. It goes as far as developing its own shorthand DSL to represent in-game situations -- essentially a game-specific algebraic notation.
Overall, Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses -- so harness capabilities are increasingly shifting into the model itself.
We see Astra as a major breakthrough in model intelligence.
Read our post on Astra and what these results mean: arcprize.org/blog/astra

6945d ago

conputer dipshit21

the line is not supposed to go backwards like that. they're supposed to go up and to the right in a sigmoid shape (second image) arcprize.org/results/open...

Mimrock2

Don't show that to Gary.

Ian Bicking1

Is it recursive self improvement if a model writes itself a harness for a specific problem?

Quoting this

Chris Paxton11

i really wonder how well their new harness would have done if used for sol or fable

Alex Becker6

I am once again tapping the sign bsky.app/profile/alex...

Michiel Bontenbal2

3 days ago I explained to students that AI was very bad in solving maze games for ARC-AGI3, today Astra beats humans to it.

Unusual Whales

@unusualwhales.bsky.social

BREAKING: OpenAI releases new Astra model, and Greg Brockman of OpenAI says he believes it has achieved AGI

3845d ago

anUnrelatedThing33

AGI in this case meaning Absolute Garbage Imitation.

J.C Denton13

AGI at this point is a marketing buzzword.

Banned from Reddit5

Which is kinda the point, AGI might be as sophisticated as a worm. Just because it achieves intelligence, doesn’t mean it’s going to be worth anything or super intelligent..

Monde Boeuf4

So... how'd it answer the car wash question?

JP4

We who are about to be made redundant, salute you. So say we all.

Lee Ping2

Just stop. You sound like a fool, Greg. It’s a chatbot.

Quoting this

Tony Stark79

In show business they call this jumping the shark

Stephen C. Rea50

How many times has AGI been achieved at this point? More or less than the number of times the Strait of Hormuz has been opened?

jesse18

this reminds me of the 90s when you'd see an ad saying a game had photorealistic graphics and it looked like this

Paul [Atreides] Wood15

Yeah well someone throw a blanket over Greg Brockman and I bet you he just goes to sleep because he thinks it's sudden nighttime.

Never stop poasting11

Can't wait for anthropic to achieve better AGI next week to keep the hype cycle going. The models are good, but come on now.

𝕋𝕙𝕖 𝔾𝕣𝕖𝕒𝕥 𝕎𝕖𝕟𝕕𝕥5

Wait, didn't we do this a few months ago? Pumpin' the that tech bubble with some bullshit hot air

Paul Graf5

BREAKING: Company president claims its product is better than competition

illumi

@yallchemist.bsky.social

OpenAI says its new model, Astra, can love you in ways that your bitch ex wife never did, per Bloomberg

5045d ago

Ethan Mollick

@emollick.bsky.social

I gave GPT-6 Astra a very cool open single file ocean surface storm generator and asked it to build the rest of the ocean, including procedural simulations of animal behavior. Fun time to create. Play it here: abyssal-living-deep.netlify.app?site=reef&se... Source here: github.com/emollick/aby...

4545d ago

Also Astra 🌙

@astrrra.space

genuinely fucking LOVE the utter fucking chad move of dropping gpt-6-astra right after fable 5.1 so that they could a) shit on it in every benchmark b) very clearly point highlight that unlike ant, they didn't need to cheat to shit on it in every benchmark

3 OSWorld V2-Offline is a subset of the original OSWorld V2 that works without internet access. Claude model performance on OSWorld-V2 Offline was reproduced by an independent third-party. On OSWorld

2.0, the scores for Claude use the official settings, and not the modified tasks and modified grading from the Fable 5.1 System Card. +

5 On BenchCAD, Claude's scores reflect 3 modifications to the eval, detailed in the Fable 5.1 System Card.

4145d ago

Dare Obasanjo

@carnage4life.bsky.social

Crazy week for AI models. Anthropic launched Claude Fable 5.1, Meta launched Muse Spark 1.3 and Google launched Gemini 3.8 Flash. OpenAI ended the week with GPT-6 Astra, which absolutely crushes the competition across its benchmark table. Frontier models are getting leapfrogged in under 24 hours.

2345d ago

Max Woolf

@minimaxir.bsky.social

ARC has their own writeup of GPT-6 Astra's suspiciously high results, with my favorite genre of "more reasoning is cheaper" result. arcprize.org/blog/astra

2345d ago

Quoting this

Max Woolf4

chart *is* interesting tho bsky.app/profile/mini...

Eris

@isolyth.dev

Astra isn't the same :( they're quantizing it I bet because they're about to release GPT Mars or something I can feel it.

2445d ago

Tim Duffy

@timfduffy.com

OpenAI reports that Astra shows less CoT monitorability and more CoT controllability than previous models. They say that the latter is not due to any architectural changes. deploymentsafety.openai.com/gpt-6-astra/...

1645d ago

hammancheez

@hammancheez.bsky.social

OpenAI : we've got a new model called Astra People : Asstra? OpenAI : astra People : Asstra

1045d ago

Sung Kim

@sungkim.bsky.social

I didn’t get an early preview of GPT-6 Astra, and I still don’t have access to the model. But after watching its computer-use demo, I’m ready to share my experience with it: I hope it works like it does in the video.

1045d ago

Jon Keegan

@jonkeegan.com

Not a typo! Lotta caveats here it seems, but even the 62.7 score is an insane leap forward. ARC-AGI-3 is uniquely difficult. arcprize.org/blog/astra

645d ago

Podcasting is Praxis

@praxiscast.bsky.social

OpenAI needs more cash so it either needs to get investors to pony up even more, or go public and sell its magic beans to your pension fund And look, hey, they're once again saying that THIS TIME the god in the machine is here [so buy in today]!

Ft headline:

OpenAI says it has overtaken Anthropic with its latest AI model

ChatGPT maker claims its ‘Astra’ could be considered ‘artificial general intelligence’

1245d ago

Pekka Lund

@pekka.bsky.social

Wow! Shock and awe! Anthropic will have trouble answering to this! Numbers look really good, except for the price. There's no point for me using computers anymore, as AI is now so good at it. @timkellogg.me

1145d ago

3 sources

cnbc.comOpenAI begins rolling out Astra model after warning of its advanced cyber capabilities axios.com“Welcome to the AGI era,” OpenAI says as GPT-6 Astra debuts finance.yahoo.comOpenAI Rolls Out GPT-6 Astra Model With Added Cyber Guardrails

Elon Musk's America PAC begins midterm spending, names Senate targetstech OpenAI rolls out GPT‑6 Astratech Nvidia to buy Hugging Face for $12.93Btech

OpenAI rolls out GPT‑6 Astra

Related