Released / limited rollout
GPT-6 Astra is here, and with it, OpenAI claims to have achieved AGI.
OpenAI president Greg Brockman says he believes the company has reached AGI. Astra dominates the reported cyber tests, trails Fable 5.1 on some broader measures, and still works like an LLM. As far as we are concerned, the AGI claim is hype.
SEP 03
Released in limited rollout
1.05M
Context window
128K
Maximum output
$10 / $50
Input / output per 1M tokens
What changed
Astra shipped. The evidence changed with it.
The model is real, priced, documented, and entering production. Most scores below still come from OpenAI's launch evaluation, so the harness matters as much as the number.
OpenAI-reported launch results, except Independent
Professional
41.4%
AutomationBench
Compared with Fable 5.1 31.4%
Computer use
72.6%
OSWorld 2.0
Compared with Opus 5 70.2%
Coding
57.9%
Terminal-Bench 4.0
Compared with Fable 5.1 55.8%
Science
97.6%
FrontierMath Tier 4
Compared with Fable 5.1 87.8%
Cyber
100%
ExploitBench
Compared with Sol 78.5%
Alignment
0%*
ExploitGym honeypot
Compared with Sol 48.2%*
Independent aggregate
61.2
AA Intelligence Index
Compared with Fable 5.1 65.7
Independent coding
67.0
AA Coding Agent Index
Compared with Opus 5 68.1
* Lower is better. Launch scores are maximum at any reasoning effort and may differ from production behavior. The two Artificial Analysis rows are the useful counterweight: Astra is not first on every broad or coding aggregate.
Dreamers' view / the AGI claim
Start by getting all basic tasks right, 100% of the time.
Astra dominates the reported cyber tests. On the Artificial Analysis Intelligence Index shown at launch, it scores 61.2 against Fable 5.1's 65.7. Its behavior still looks like an LLM to us: remarkable answers alongside tasks that need correction. Calling that AGI feels like hype.
OpenAI reports 41.4% on AutomationBench, 72.6% on OSWorld 2.0, and 57.9% on Terminal-Bench 4.0. Strong gains. Plenty left to get right. The 100% standard above is our demand for dependable execution; human intelligence also makes mistakes. Launch evaluation.
Availability
Released does not mean broadly available yet.
Astra launched to a limited set of organizations. OpenAI says Plus, Pro, Business, Enterprise, API, Azure, and AWS Bedrock access will follow over the coming days.
Enterprise access is off by default. The API model ID is gpt-6-astra. Pro, Business, and Enterprise will also receive Astra Pro.
The consequential result
The first OpenAI model rated Critical for cybersecurity.
OpenAI says Astra can find previously unknown flaws and develop exploit chains against hardened systems with limited human guidance. In testing, it found two zero-days and reached 100% on ExploitBench.
The launch model refuses advanced exploit requests. Less restrictive defensive access begins with a small alpha group and is expected to expand through Daybreak.
The system card rates Astra High, not Critical, for biological and chemical capability, and below OpenAI's High threshold for AI self-improvement.
The Cursor question
Astra is shipping around Cursor.
OpenAI has proposed ending model supply to Cursor on November 12 after its acquisition by SpaceX. The stated issue is confidence in contract and terms-of-service compliance.
Astra's listed channels are OpenAI products, its API, Azure, and AWS Bedrock. For a model with Critical cyber capability, identity, monitoring, authorization, and the operating environment are now part of the product.
Evidence boundary
What remains unknown
A system card is evidence. A launch table is evidence. Neither substitutes for many users running the production model on comparable work.
- 01Independent replication across more public harnesses
- 02Cost per completed task at each reasoning effort
- 03Production reliability, refusal friction, and tool-call failure rates
- 04Which advanced cyber workflows move from alpha access into Daybreak
Short answers
GPT-6 Astra FAQ
When is Astra coming out?
GPT-6 Astra launched on September 3, 2026 to a limited set of organizations. OpenAI says access for ChatGPT Plus, Pro, Business, and Enterprise users, the OpenAI API, Microsoft Azure, and AWS Bedrock will roll out over the following days.
What is GPT-6 Astra?
GPT-6 Astra is OpenAI's newest frontier model for computer use, professional work, coding, research, and long-running agent tasks. OpenAI released it on September 3, 2026 to a limited set of organizations, with broader access rolling out over the following days.
How much does GPT-6 Astra cost?
Standard API pricing is $10 per million input tokens, $1 per million cached input tokens, $12.50 per million cache-write tokens, and $50 per million output tokens. Requests above 272K input tokens carry long-context multipliers. Task cost still depends on token use, tools, retries, and completion time.
Is GPT-6 Astra the best model?
Not on every measure. OpenAI reports category-leading results on computer use, professional automation, Terminal-Bench 4.0, FrontierMath Tier 4, and cyber evaluations. Its own table shows Claude Fable 5.1 ahead on the Artificial Analysis Intelligence Index and Claude Opus 5 narrowly ahead on the Artificial Analysis Coding Agent Index.
Will GPT-6 Astra be available in Cursor?
OpenAI says it will stop providing models to Cursor after Cursor's acquisition by SpaceX, with a proposed November 12, 2026 cutoff. Astra is instead rolling out through OpenAI products, the API, Microsoft Azure, and AWS Bedrock.
Source ledger