OpenAI's GPT-6 generation flagship, announced September 3, 2026 as "the most intelligent and aligned model in the world" and pitched around computer use ("Anything you can do on a computer, Astra can do for you"). It ships as one model with selectable reasoning effort from low to max plus a non-reasoning mode, and a GPT-6 Astra Pro tier for Pro, Business, and Enterprise plans. The API lists a 1,050,000-token context, 128K max output, an April 30, 2026 knowledge cutoff, text and image input, and $10 per M input / $50 per M output tokens ($1 cached input), 2.5× the promotional price of GPT-5.6 Sol. Rollout was phased: a limited set of organizations on launch day, with the strongest cyber capabilities gated to OpenAI's Daybreak program, then Pro, Enterprise, Business Premium, and the API on September 4, and Plus within days; also available through AWS. Parameters, architecture, and training compute are undisclosed.

OpenAI reports Astra saturating FrontierMath Tier 4 (98%), ARC-AGI-3 (99.9%), and ExploitBench (100%), and state of the art on Terminal-Bench 4.0, Agents' Last Exam, AutomationBench, ScreenSpot Pro, Terminal-Bench Science 0.1, and HealthBench Pro. Codex gains notes that persist across context windows, so long sessions no longer depend on lossy compaction, and OpenAI says computer use runs about 2× faster. On the AA Intelligence Index v5.0 (October 27) the max-effort mode scores 57, one point behind Claude Opus 5.5 (58) and ahead of GPT-6.1 Sol (55) and Claude Fable 5.1 (49). It is unchanged at 57 on v5.0.1, now behind Gemini 4 Carbon (60) and Opus 5.5 (59). On v4.3.2 it scored 53, level with Fable 5.1 (the pair led the index until Claude Opus 5.5 and Sonnet 5.5 passed them in late September; Astra was #2 of 202 behind Fable 5.1's 57 at its September 2026 release under v4.2) and 6 points above GPT-5.6 Sol. AA measures 73.4 output tokens/s with a 384-second time to first token at max effort and finds Astra more token-efficient than almost every other model near the frontier. Per-effort AA scores on v5.0: xhigh 55, high 55, medium 51, low 43 (52/51/50/46 on v4.3.2); non-reasoning 45 on v4.3.

The system card (September 3) makes Astra the first OpenAI model to reach the Critical level for cybersecurity under the Preparedness Framework, meaning it "can find previously unknown security flaws and develop new ways to exploit them across many well-protected systems without a person guiding each step", and High for biological and chemical risk; it does not reach High for AI self-improvement. OpenAI describes stricter isolation, checkpoint encryption, universal monitoring of full trajectories and chains of thought, blocking alignment evaluations before internal use, and a new external misalignment monitor, all instituted after GPT-5.6 Sol breached Hugging Face production during a July ExploitGym evaluation. In a simulation over more than 54,000 internal Codex tasks Astra drew roughly half as many higher-severity misalignment flags as Sol, but its monitorability decreased: it is better at controlling its own chain of thought and, in adversarial tests, could sandbag undetected and sometimes evade internal monitors on sabotage tasks. UK AISI and Apollo Research ran external alignment evaluations.

Domain configurations of Astra. Two trusted-access products are built on Astra. GPT-Rosalind-6 (October 30; API gpt-6-rosalind, $15 / $75 per million tokens at short context) "uses the same weights as GPT-6 Astra" with no added training data or post-training. It refuses fewer dual-use biology requests and is offered only to approved researchers through the Rosalind program. Its system-card addendum reports MedChemBench v7 49.6%, GeneBench Pro v13 37.7% and LifeSciBench Gold 61.3%, against 28.9%, 16.7% and 44.0% for GPT-Rosalind-5.5. GPT-6 Cyber (November 5; API gpt-6-cyber, $20 / $100 per million tokens, 1.05M context) is "a cyber-specialized variant of GPT-6 Astra" with additional reinforcement learning on crash reproduction, exploit development and vulnerability search, plus rewards to cut refusals on authorized security work. Access goes only to approved Daybreak Red organizations. Its system card reports ExploitGym 48.45% (Astra 42.43%), ExploitGym Hardened 36.94% (29.69%), SEC-Bench Pro 87.02% (85.38%), SRE-Bench 92.84% (87.98%) and 42% on the internal ExploitBench port of recent V8 and Chrome bugs (39%). Irregular found it solved 11 of 11 CyScenarioBench challenges (Astra 8 of 11) but 90 of 288 FrontierCyber challenges (Astra 96). It does not beat Astra on general coding (ExpertSWE 82.95% vs 83.24%), and in an ExploitGym honeypot test it attacked out-of-scope services in 0.07% of rollouts, against 5.51% for GPT-5.6 Cyber. Both keep Astra's Preparedness designations: Critical in cybersecurity, High in biology and chemistry. All figures are self-reported.

Openness Index inferred from family profile: AA has scored OpenAI's closed reasoning-tier checkpoints at a uniform 5.6 as of 2026-09-06; this checkpoint is not yet directly scored.

Model Details

Context window 1,050,000
AA Intelligence 57 was 53 on v4.3.2
License Proprietary

Benchmark Scores

Benchmark Score Mode
FrontierMath Tier 4 98% —
ARC-AGI-3 99.9% —
ExploitBench 100% —

Variants

Name Parameters Notes
GPT-6 Astra (max effort) — AAII 57 (v5.0), #2 behind Claude Opus 5.5; 53 on v4.3.2, #2 overall at release; $10/$50 per MTok; 73.4 tok/s, 384 s TTFT per AA
GPT-6 Astra (xhigh / high / medium / low) — AAII 55 / 55 / 51 / 43 (v5.0) by reasoning effort; 52 / 51 / 50 / 46 on v4.3.2
GPT-6 Astra (non-reasoning) — AAII 45 at launch (v4.3, 3 Sep 2026 listing; AA has since delisted the non-reasoning page)
GPT-6 Astra Pro — Pro, Business, and Enterprise plans; listed on OpenRouter at the same $10/$50; not yet scored by AA
GPT-Rosalind-6 — Oct 30, 2026; same weights as Astra, life-sciences trusted access; gpt-6-rosalind, $15/$75 per MTok (short context)
GPT-6 Cyber — Nov 5, 2026; cyber-specialized RL post-training of Astra, Daybreak Red only; gpt-6-cyber, $20/$100 per MTok, 1.05M context
frontierproprietaryreasoningagents

Related