OpenAI dropped GPT-6 Astra today, and the positioning is a clear pivot: this is a computer-use model first, a chat model second. The pitch is that Astra operates browsers, spreadsheets, terminals, and desktop apps the way a person would — executing multi-step jobs rather than explaining them. The headline number is 72.6% on OSWorld V2-Offline, up from GPT-5.6 Sol’s 65.7%, with average task time nearly halved from 75 minutes to 40. But don’t expect to spin it up on your own cluster. Astra is closed and hosted only, with no released weights, and it’s live today exclusively for organizations in OpenAI’s Trusted Access and Daybreak programs.
For developers, the meaningful change is context handling. Codex previously used compaction — summarizing earlier turns when context filled up — which quietly discarded the details an agent later needs: why a fix failed, which tests ran, what requirement was added at the start. Astra replaces that with persistent notes across context windows and search into earlier messages and tool output. The feature ships experimental behind a config.toml flag and becomes the Codex default in the coming weeks. Astra can also ask the user a clarifying question while continuing work that doesn’t depend on the answer, addressing a classic failure mode where one unresolved decision stalls an entire job.
The benchmark picture is more mixed than the launch framing suggests. Astra scores 99.9% on ARC-AGI-3, but that number came from a Responses API harness with reasoning retained between turns and compaction for long contexts — OpenAI has previously shown those settings move ARC-AGI-3 scores substantially without changing the underlying model. So the result measures the model plus the agent system, not the model alone. On coding, the gains are marginal: 74.1% on DeepSWE v1.1 versus 72.7% for Sol, while Meta’s Muse Spark 1.3 sits at 75.4% and Gemini 3.8 Flash and Claude Opus 5 hover near 74%. On a 113-task benchmark, that’s a gap of one or two tasks. The FrontierMath Tier 4 score of 97.6% comes with a caveat too: Epoch AI notes OpenAI funded FrontierMath and has exclusive access to part of it.
Then there’s the cybersecurity angle. Astra is the first model OpenAI has designated as reaching the Critical threshold in its Preparedness Framework. In testing, it developed exploits for hardened browsers and operating systems and found two previously unknown V8 vulnerabilities now being disclosed to maintainers. The practical consequence: standard access refuses advanced cybersecurity work, and an API safety check stops tasks outright rather than pausing for approval. OpenAI’s Mia Glaese warned that users outside trusted-access programs may hit slowdowns, pauses, or blocks — sometimes during unrelated work. Pricing lands at $10 per million input tokens and $50 per million output, with cached input at $1.00 and a 2x input surcharge above 272K tokens. That’s a premium worth paying only if the agent actually finishes the job, and on coding, the data says it’s still in the pack rather than ahead of it.
