Skip to content 
GPT-6 Astra launched on September 3, 2026, and Sam Altman told CNBC it marks a new capability level. OpenAI’s own scores are near-perfect on a hard math exam, an abstract reasoning test, and a hacking benchmark, but outside evaluations are more mixed. This article separates what independent testing has confirmed from launch-day marketing.
What OpenAI Says GPT-6 Astra Can Do?
OpenAI calls Astra the world’s most intelligent and aligned model, strongest at computer use, coding, science, security, and professional work. Its launch post says Astra finishes computer tasks in about half the time GPT-5.6 Sol needs while scoring higher. On an alignment test, Sol overstepped its authorization 48% of the time without safeguards, while Astra never did. All figures are OpenAI’s own.
The ARC-AGI-3 Result, Explained Properly
What ARC-AGI-3 actually tests?
An AI enters an unfamiliar game with no instructions, works out the rules, finds the goal, and plans its way there.
Two scores, two setups
Astra scored 62.7% under ARC Prize’s neutral harness, the same minimal setup every model gets. It scored 99.9% under OpenAI’s Provider Adapter, which lets the model keep its reasoning between requests. Both are records, but the neutral score is the fairer comparison across vendors.
The finding that matters more than the score
ARC Prize compared Astra with about 500 volunteers. In the Provider Adapter run, it used fewer moves than the typical person on 96% of levels. ARC Prize had expected humans to hold that edge longer.
What ARC Prize is careful to say?
The foundation does not claim AGI. Its games follow closed, fixed rules, which real life doesn’t.
The Behaviour Researchers Found Most Interesting
It invented its own shorthand
Astra kept compact, code-like notes on objects, coordinates, and plans. ARC Prize calls it an on-the-fly notation, not a full language, and says Astra’s was unusually precise.
It wrote its own software
In a separate coding sandbox, Astra built helper programs mid-game, such as path solvers. Human testers had no such tools, so this shows the model plus its tools, not a fair human comparison.
What Early Testers Reported?
As reported by Marketing4eCommerce, Carlos Santana praised the benchmarks without having access, while Allie K. Miller, who did, called computer and browser use excellent but said Astra failed her wordplay challenges and underwhelmed on 3D and animation. She’d still use Claude Fable 5.1 for large architectural builds. Tom Krcha rebuilt a 3D house from one photo. Playco, in an OpenAI-published case study, reported half the manual fixes of the prior model.
The Cybersecurity Milestone
OpenAI says Astra reaches the Critical cyber level under its Preparedness Framework. It reports a perfect ExploitBench score, 88% on a reverse-engineering test versus 56% for Sol, and two previously unknown vulnerabilities found in testing. We found no independent replication so far. The public version refuses advanced offensive tasks.
Where Independent Benchmarks Disagree?
The efficiency story is better than the intelligence story
Artificial Analysis scored Astra 53 on its Intelligence Index v4.3, tied with Claude Fable 5.1 and six points above Sol. The bigger news is efficiency: GPT-6 Astra matched Fable 5.1 at about two-fifths of the cost per task, though it costs about 60% more than Sol. Epoch AI’s composite ranked it first at launch, per The Decoder.
What got better and what got worse
Hallucinations on Artificial Analysis’s knowledge test fell from 92% to 51%, and coding and workflow scores rose. Occupational tasks and slide design slipped, both areas OpenAI promotes. Earlier coverage used an older index showing Astra level with Sol, and OpenAI edited five launch figures after publishing, per TNW, so recheck numbers.
What This Means If You Run a Business?
This part is our interpretation. The edge is agent work, not raw intelligence. Multi-step jobs like form-filling and coding show the clearest gains, while general reasoning is a tie with Fable 5.1, not a lead.z
Cost per task beats headline score. Tokens cost 2.5 times Sol’s, but per-task costs run from about 15% higher on coding to about 60% higher on general reasoning. Test your own workload.
Watch the security shift. Finding zero-days is a real milestone, access is tightly controlled, and safety checks can pause legitimate work.
When Can You Use It?
As of September 19, 2026, OpenAI describes Astra as available in ChatGPT Work, Codex, and the API, and names Azure and AWS Bedrock as channels. API pricing is $10 per million input tokens and $50 per million output tokens.
Where DigiEvolve Fits?
DigiEvolve haven’t tested Astra ourselves, so we won’t call it a guaranteed win. We can help you work out where agentic AI fits your operations, using demonstrated results instead of launch-day claims.
Frequently Asked Questions
What is GPT-6 Astra?
OpenAI’s newest model, launched September 3, 2026, aimed at computer use, coding, science, and security.
What is ARC-AGI-3 and why does it matter?
A benchmark where an AI must learn unfamiliar games without instructions, meant to gauge progress toward general intelligence.
Did GPT-6 Astra really score 99.9%?
Yes, under OpenAI’s Provider Adapter. Under ARC Prize’s neutral harness it scored 62.7%.
Is GPT-6 Astra better than Claude or Gemini?
It ties Claude Fable 5.1 on Artificial Analysis’s index at lower cost per task. OpenAI’s table only includes Gemini 3.8 Flash, so there’s no like-for-like Gemini comparison.
Is GPT-6 Astra AGI?
No. ARC Prize says it isn’t claiming AGI.
How much does GPT-6 Astra cost compared to GPT-5.6?
It’s 2.5 times Sol’s rates: $10 and $50 per million tokens versus $4 and $20.
When will GPT-6 Astra be available to everyone?
OpenAI says it’s available in ChatGPT Work, Codex, and the API. We found no confirmation of free access.
