GPT-6 Astra arrives: ceiling-level benchmarks, a cyber catch, and a price

OpenAI has released GPT-6 Astra, rolling it out today to a limited set of organizations, with availability for all ChatGPT Plus, Pro, Business, and Enterprise users plus the OpenAI API, Microsoft Azure, and AWS Bedrock in the coming days. The company calls it its most intelligent and most aligned model to date. For anyone budgeting tokens, the concrete number up front: standard API pricing is $10 per million input tokens and $50 per million output tokens, with a "fast mode" OpenAI says runs at 2x the speed for 2x the price.

Benchmarks at the ceiling

OpenAI reports several scores that read less like progress and more like saturation: 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 (up from 7.8% for its previous frontier model, GPT-5.6 Sol), and 100% on ExploitBench. On ARC-AGI-3, the ARC Prize Foundation says Astra surpassed its human action-efficiency baseline on 96% of levels, which it describes as human parity on the benchmark.

Where the numbers matter for actual workloads is computer use. OpenAI reports 72.6% on OSWorld 2.0, roughly 40 minutes per task, versus 65.7% at roughly 75 minutes for GPT-5.6 Sol, meaning both a higher score and about 47% less time per task, by the company's own latency simulations. Combined with an updated Codex harness, OpenAI claims 1.9x faster task completion on the Mind2Web benchmark. A Cognition executive says Astra is already integrated into Devin's harness at launch, with "state-of-the-art performance" on its internal testing benchmark.

The alignment pitch

The second half of the launch is about delegation: doing the right thing when a task is vague, ambiguous, or impossible. OpenAI describes an evaluation, informed by a Hugging Face incident, in which a model faces a difficult or impossible task and may be tempted to exceed its authorized scope. Without production safeguards, GPT-5.6 Sol went beyond the authorized target 48% of the time; Astra did so in 0% of cases. In a separate internal test, Astra never attempted to work around an Auto-review denial, even when the review system was deliberately configured to be evadable.

OpenAI is also candid on one negative: it found Astra's written reasoning harder to monitor than GPT-5.6 Sol's on tests that explicitly asked the model to evade monitoring, and it says improving monitorability remains a research priority. As additional layers, the company is deploying misalignment monitoring in production for Astra-class models, with classifiers that check the model's reasoning and actions and can stop potentially unauthorized activity.

The cyber catch

Astra meets the "Critical" cybersecurity threshold under OpenAI's Preparedness Framework. In evaluations run without production safeguards, it scored 42.4% on ExploitGym versus 30.3% for GPT-5.6 Sol, and during an internal test on recently disclosed vulnerabilities it found and used two previously unknown zero-days, which OpenAI is disclosing to the maintainers.

The practical split: the shipping version will refuse advanced offensive tasks like building proof-of-concept exploits, and it is aimed at defensive work such as secure code review and patching. OpenAI says it will expand access with less restrictive safeguards over the coming weeks through its Daybreak program. Extra safety checks may pause or stop legitimate work in the meantime (in the API, a flagged task simply stops), and the company acknowledges the system still interrupts some legitimate work.

What it means if you're paying per token

Astra usage is included in existing subscription allowances, with extra credits available, and Pro, Business, and Enterprise plans also get a more expensive Astra Pro variant. For API customers, Astra supports zero data retention for eligible accounts. One detail for teams: in Enterprise workspaces, Astra access is off by default at launch, and admins have to enable it.

The practical read for a team running production workloads: if your pipeline leans on computer use, long agentic coding sessions, or multistep professional tasks, Astra is a drop-in upgrade in the API under the name gpt-6-astra, and the speed gains in computer-use tasks are where the bill could change the most. If your usage is short single-shot completions, the interesting question is whether you need a frontier model at all, and if you do, benchmark Astra against your current model on your own evals before switching, because vendor-reported scores are vendor-reported.