OpenAI Releases GPT-6 Astra: A 1.05M-Context Computer-Use Model Gated Behind a ‘Critical’ Cyber Threshold

Details

OpenAI released GPT-6 Astra today, positioning it as a computer-use system rather than a chat model. The defining architectural change is persistent context: instead of compacting earlier turns into summaries (which discards details like why a fix failed or which requirement was added), Astra keeps notes across context windows and can search back through earlier messages and tool output. Two new reasoning.effort tiers — xhigh and max — extend beyond the previous high ceiling. The model also supports parallel clarification: asking the user a question while continuing work that does not depend on the answer. Benchmark numbers are strong but carry caveats. Astra scores 72.6% on OSWorld V2-Offline versus 65.7% for GPT-5.6 Sol, cutting average task time from 75 to 40 minutes. Anthropic reports 77.9% for Claude Fable 5.1 on a different OSWorld release — direct comparison is unreliable. On ARC-AGI-3, Astra scores 98.6% using a Responses API harness with persistent between-turn reasoning; OpenAI has previously shown this harness configuration moves scores substantially, so the number measures the system, not the bare model. Epoch AI notes OpenAI funded FrontierMath and holds exclusive access to part of it. Coding is the relative weak spot: 74.1% on DeepSWE v1.1, behind Meta's 75.4% for Muse Spark 1.3 and near Claude Opus 5 and Gemini 3.8 Flash on the public leaderboard. Model specs: 1,050,000-token context, 128,000 max output tokens, April 30 2026 knowledge cutoff, text+image input, text-only output. No fine-tuning. Hosted-only, no weights released. Deployment is gated. Astra is live only for organizations in OpenAI's Trusted Access and Daybreak programs at $10 input / $50 output per million tokens. General availability is not announced. For practitioners building autonomous agents, the persistent-context architecture is the most practically significant change — whether real pipelines stall less will be the real-world test.