GPT-6 Astra
GPT-6 Astra, OpenAI's first model rated Critical for cybersecurity, claims 97.6% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3.
Press accounts say it was pretrained on 100,000+ GPUs at Stargate Texas with a looped "recurrent depth" design; RL for computer use, coding and science. Rollout was phased and trust-gated, and it refuses advanced exploit work until Daybreak access. Its written reasoning is harder to monitor than GPT-5.6 Sol's.
- Date
- Thursday, 3 September 2026
- Lab
- OpenAI
- Kind
- model
- Access
- closed API
- Price
- $10 input / $50 output per 1M tokens; Fast mode 2.5x speed at 2x price
Figures
| Measure | Value | Measured by |
|---|---|---|
| FrontierMath Tier 4 (v2) | 97.6% GPT-5.6 Sol 83.0%; company-run | company |
| ARC-AGI-3 | 99.9% with OpenAI harness 62.7% in ARC default harness per Simon Willison; GPT-5.6 Sol 7.8% | company |
| Terminal-Bench 4.0 / DeepSWE v1.1 | 57.9% / 74.1% GPT-5.6 Sol 37.3% / 72.7% | company |
| OSWorld 2.0 | 72.6% GPT-5.6 Sol 65.7% | company |
| ExploitBench / SRE-Bench (no production safeguards) | 100% / 88.0% GPT-5.6 Sol 78.5% / 55.9% | company |
| Artificial Analysis Intelligence Index v4.1.1 | 61.2 GPT-5.6 Sol 60.9; Claude Fable 5.1 65.7; as listed in OpenAI's table | independent |
This is contested. OpenAI calls Astra its most aligned model; critics note lower monitorability and higher verbal eval awareness (9.6% vs 2.8% for Sol per Zvi Mowshowitz reading the system card; UK AISI and Apollo flagged confounds). The looped-transformer and 100,000-GPU details come from Wikipedia summarizing press. OpenAI's page does not give them.
Sources
- openai.com/index/gpt-6-astra/
- deploymentsafety.openai.com/gpt-6-astra/
- simonwillison.net/2026/Sep/3/gpt6-astra/
- thezvi.wordpress.com/2026/09/09/gpt-6-astra-the-system-card-alignment-and-what-comes-next/
- en.wikipedia.org/wiki/GPT-6
- developers.openai.com/api/docs/models/gpt-6-astra
This record was checked and corrected against its sources on 6 October 2026. How we check