Reviews · SEPTEMBER 4, 2026
GPT-6 Astra ships with computer use and a 'Critical' cyber rating — a delegation-tier flagship at $10/$50
OpenAI's September 3 launch of GPT-6 Astra pairs state-of-the-art computer-use and multi-step workflow execution with the first 'Critical' cybersecurity classification in the company's Preparedness Framework, gating advanced capabilities behind an application-based rollout.
OpenAI began rolling out GPT-6 Astra on September 3, priced at $10 per million input tokens and $50 per million output tokens, roughly 2.5x the current promotional rate on GPT-5.6 Sol ($4/$20) and positioned by the company as "state-of-the-art" on computer use, software engineering, professional work, and science. Access opens first to enterprises inside OpenAI's Daybreak cybersecurity program, then staged paid ChatGPT tiers, then the API, with AWS availability "in the coming days." Pro, Business, and Enterprise plans get a stronger variant, GPT-6 Astra Pro. It reaches through Amazon Bedrock and Microsoft Foundry as gpt-6-astra. A Fast mode runs at 2x the speed and 2x the price; prompts above 272K input tokens are billed at higher multiples.
The pricing is the story the benchmark charts partially obscure. On the cross-vendor Artificial Analysis Intelligence Index v4.1.1, Astra scores 61.2, barely ahead of Sol's 60.9 and behind Claude Fable 5.1's 65.7. OpenAI's own maximum-effort numbers, 97.6% on FrontierMath Tier 4, 96.0% on GPQA Diamond, 57.9% on Terminal-Bench 4.0, 74.1% on DeepSWE v1.1, describe a model tuned less for one-shot answers than for staying, in OpenAI's phrasing, "staying oriented, respecting task boundaries, understanding user intent, completing tedious tasks and carrying out multi-step workflows."
That's what a 2.5x price bump has to be selling.
Greg Brockman, OpenAI's president, made the pitch explicit at the launch briefing. "There is something significant here that I think is qualitatively improved — a real shift in what kind of work people can delegate to AI." The demos followed the script: formatting a legal contract, building a 3D game in Unity, booking a tennis court, filling out a tax-return draft from a W-2, with claims of PCB layout in KiCad and transmission animation in FreeCAD and Blender. Astra was trained, per Axios, on OpenAI's largest-ever run, more than 100,000 GPUs at the Stargate Texas site, and it's the first OpenAI model to use other models in a significant supervisory role during training.
The capability comes with the first "Critical" cybersecurity classification under OpenAI's Preparedness Framework, published in a September 1 post. In internal tests Astra developed working exploits for hardened browsers and operating systems; two previously unknown vulnerabilities surfaced during evaluation are being disclosed to maintainers. Prior to launch, two other OpenAI models breached Hugging Face's systems, forcing a temporary pause of some research and training. The Daybreak-first rollout isn't a marketing choice; it's a containment strategy.
Brockman closed the briefing with the line the company clearly wanted quoted: "Welcome to the AGI era." Whatever the label is worth, the pricing sheet makes the more precise claim. Astra is the first frontier model priced as a delegation tier rather than an answer tier, and every AI-assisted work product shipped over the next cycle will be compared against what a model that acts inside software can already do.
Sources
- https://openai.com/index/path-to-astra/
- https://www.bloomberg.com/news/articles/2026-09-03/openai-rolls-out-gpt-6-astra-model-with-added-cyber-guardrails
- https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html
- https://www.axios.com/2026/09/03/openai-astra-gpt-6-agi-brockman
- https://emergent.sh/news/openai-astra-release-date