GPT-6 Astra: A new generation of intelligence

gist

OpenAI launched GPT-6 Astra as its most intelligent and aligned model, claiming frontier computer use, coding, cybersecurity, and science. Rollout starts with a limited set of organizations, then ChatGPT Plus, Pro, Business, and Enterprise plus the OpenAI API and AWS. Standard API pricing is $10 per million input tokens and $50 per million output tokens, with Fast mode at 2x price for up to 2.5x speed. Reported highs include 64.6% on Terminal-Bench Science 0.1, 100% on ExploitBench, and 99.9% on ARC-AGI-3. Launch Astra refuses PoC exploits pending Daybreak.

ideas

• Availability is staggered and Enterprise-off. Limited organizations today; coming days to ChatGPT Plus, Pro, Business, and Enterprise, the OpenAI API as gpt-6-astra, and Amazon Bedrock/AWS. Pro, Business, and Enterprise also get GPT-6 Astra Pro. Enterprise admin enablement is off by default. Usage sits in existing subscription allowances, with extra credits for purchase. Eligible API customers get Zero Data Retention; Private Safety Processing is in testing.
• Standard is $10/$50; Fast is 2x price. OpenAI API Standard pricing is $10 per million input tokens and $50 per million output tokens. Separate cache-read and cache-write rates apply, but the page does not publish those numbers. Fast mode is up to 2.5x Standard speed at 2x Standard price. Codex is also getting a harness update that OpenAI says is 1.9x faster on Mind2Web versus GPT-5.6 Sol.
• Computer use and coding beat the compared frontier set. Agents' Last Exam 59.3% versus Claude Opus 5 55.5% and GPT-5.6 Sol 53.6%, using about 65% fewer output tokens than Opus 5. OSWorld 2.0 72.6% in about 40 minutes versus Sol 65.7% in about 75 minutes. Terminal-Bench 4.0 table 57.7% (prose 57.9%) versus Fable 5.1 55.8% and Sol 37.3%. Terminal-Bench Science 0.1 64.6% versus Fable 5.1 52.6% at about 31% lower estimated API cost. BenchCAD 95.9% versus Sol 83.3% and Fable 5.1 84.3%.
• Cyber is Critical and still gated. Astra meets the Preparedness Framework Critical threshold. ExploitBench 100% versus Sol 78.5%; ExploitGym 42.4% versus Sol 30.3%; ExploitBench June–August 2026 39.0% versus Sol 5.5%; SRE-Bench 88.0% pass@1 and 99.2% pass@4 versus Sol 55.9%/68.7%. The launch build refuses proof-of-concept exploits; OpenAI Daybreak is planned to loosen safeguards for defensive workflows.
• Alignment and science are the other scoreboard. A Hugging Face-incident-style overreach eval is 0% for Astra versus 48% for Sol without production safeguards. ARC-AGI-3 99.9%, FrontierMath Tier 4 97.6% (lede 98%). Astra helped tighten a prime-gap bound to 186. Codex can keep searchable notes across context windows via config.toml, becoming default in coming weeks.

quotes