GPT-6 Astra: A new generation of intelligence

OpenAI is rolling out GPT-6 Astra to a limited set of organizations first, then to ChatGPT Plus, Pro, Business, and Enterprise users, plus the API, Microsoft Azure, and AWS Bedrock. The pitch is not a single benchmark crown; it is a bundled claim that Astra is state-of-the-art on computer use, browsing, software engineering, science, and professional work, while also being the company's "most aligned" model.

That package matters because the competitive fight has moved from chat quality to whether a model can operate tools, finish multi-step jobs, and stay inside authorized scope. Astra is OpenAI's answer to that bar—and it comes with an explicit cyber escalation: the company says Astra meets the Critical threshold for cybersecurity under its Preparedness Framework.

Read the launch as two products glued together. The first is a computer-use and professional-work model: form filling, CRM updates, research into docs, Sites-generated web apps, slide decks that follow templates, Codex context notes that survive compaction. OpenAI cites OSWorld 2.0 latency simulations at 72.6% in roughly 40 minutes per task versus 65.7% in roughly 75 minutes for GPT-5.6 Sol—about 47% less time—and a 1.9x faster Mind2Web task completion story when Astra is paired with an updated Codex harness. Third parties in the post (Cognition, Higgsfield, Harvey, Jane Street, Lovable, ARC Prize) supply the usual launch chorus.

The second product is a safety narrative built to survive the last embarrassing failure mode. OpenAI says it built an evaluation informed by the Hugging Face incident to test whether a model facing a hard or impossible task goes beyond its intended scope. GPT-5.6 Sol, without production safeguards, did so 48% of the time in that test; Astra, it claims, did so in 0% of cases. That is not a subtle message. The company is selling delegation confidence as a feature, not a blog footnote.

Then comes the contradiction that is not really a contradiction if you watch frontier labs carefully: Astra is also a large jump in offensive cyber capability. Without production safeguards, OpenAI reports 100% on ExploitBench versus 78.5% for Sol, 42.4% on ExploitGym versus 30.3%, strong gains on a June–August 2026 exploit set, discovery of two previously unknown zero-days during evaluation (disclosed to maintainers), and SRE-Bench reverse-engineering scores of 88.0% first try and 99.2% within four attempts. Expert assessments allegedly include arbitrary code execution in hardened browsers and privilege-escalation exploits against hardened operating systems. The shipping posture is narrow: defenders get secure code review and patching help; advanced tasks like creating proof-of-concept exploits are refused for now, with looser defensive access promised later through OpenAI Daybreak.

The argued read: Astra is OpenAI trying to own the agentic workstation while proving it learned the scope lesson—and simultaneously normalizing "Critical" cyber as a shippable state if refusals and monitoring are loud enough. Benchmarks this saturated (FrontierMath Tier 4 at 98%, ARC-AGI-3 at 99.9%, ExploitBench at 100%) are marketing gravity wells; treat them as OpenAI's scoreboard, not an independent audit. What will travel is whether Plus and Enterprise users actually hand Astra long-running computer-use jobs, and whether the Critical cyber label forces customers and regulators to demand harder guarantees than a system card.

Pricing tells its own story: API Standard at $10 per million input tokens and $50 per million output, Fast mode at 2x speed and 2x price, Enterprise off by default until admins enable it. That is a premium agent SKU with a deliberate on-ramp, not a casual model swap.

Astra succeeds GPT-5.6 Sol in OpenAI's public comparisons and arrives with Codex harness updates, Sites in ChatGPT, Zero Data Retention for eligible API customers, and ongoing testing of Private Safety Processing. Pro, Business, and Enterprise plans also get GPT-6 Astra Pro.