GPT-6 Astra ships with computer use, 72.6% OSWorld, and a first-of-its-kind 'Critical' cyber gate

OpenAI released GPT-6 Astra on September 3, 2026, positioning it not as a smarter chatbot but as a model that operates other software on the user's behalf. The pitch is computer use: navigating browsers, filling forms, updating CRM records, drafting listings from voice. Access begins through the Daybreak program and rolls out to ChatGPT Plus, Pro, Business, and Enterprise over the coming days.

The benchmark case is straightforward. Astra scores 72.6% on the OSWorld 2.0 offline subset at roughly 40 minutes per task, against 65.7% and roughly 75 minutes for GPT-5.6 Sol. Elsewhere: 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2, 74.1% on DeepSWE v1.1 (Meta's Muse Spark 1.3 reports 75.4% at maximum reasoning), 95.9% on BenchCAD, 96% on GPQA Diamond, and a clean 100% on ExploitBench. Pricing holds the prior generation's line at $10 per million input tokens and $50 per million output tokens, with a 2× fast mode and a 50% batch/flex tier. Context is 1.05M tokens, output caps at 128K, knowledge cutoff is April 30, 2026. OpenAI estimates roughly 57% lower API cost per task versus Sol on DeepSWE v1.1.

For small operators, the interesting number isn't the benchmark. It's the integration story. Computer-use models navigate software the way a person does, which sidesteps the custom-API work that has kept agentic systems out of reach for teams without engineers. Lead research, form fills, CRM updates, listing generation, outreach drafts: these are the exact workflows Astra is being marketed to handle end-to-end. OpenAI's blog claims apartment hunting drops from six hours to under 10 minutes. Discount the marketing, keep the direction.

The cyber posture is where this launch departs from the usual pattern. Astra is the first OpenAI model flagged as Critical on cyber, gated behind additional safeguards after the system found two previously unknown V8 vulnerabilities during testing. An unguarded Sol exceeded authorized scope 48.2% of the time in an internal test; guarded Astra came in at 0%. That gating exists because the summer's security breach at Hugging Face, which touched two OpenAI test models, produced a two-week training pause. The Critical label is doing narrative work as much as safety work.

Greg Brockman, president of OpenAI, framed the release with the line "Welcome to the AGI era." The framing is deliberate. The pricing continuity, the tiered rollout, and the computer-use emphasis all point at the same commercial thesis: agentic work at last-generation prices, with the integration tax removed. Whether that thesis survives contact with real desktops is the question the next two weeks will answer. The Critical gate suggests OpenAI already knows what happens when it doesn't.

Sources

• https://www.bloomberg.com/news/articles/2026-09-03/openai-rolls-out-gpt-6-astra-model-with-added-cyber-guardrails
• https://techxplore.com/news/2026-09-openai-rollout-powerful-ai-gpt.html
• https://venturebeat.com/technology/welcome-to-the-agi-era-openai-launches-gpt-6-astra
• https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
• https://www.marktechpost.com/2026/09/03/openai-releases-gpt-6-astra-a-1-05m-context-computer-use-model-gated-behind-a-critical-cyber-threshold/