Quick answer

OpenAI released GPT-6 Astra on September 3, 2026 — its biggest capability jump of the year, arriving days into the GPT-5.7/Claude Opus 4.9 exchange. The headline shift is from answering questions to completing multi-step work directly: computer use (operating a screen, filling forms, testing software), agentic coding that holds an objective across changing instructions, and stronger documents/spreadsheets/presentations output. OpenAI reports very high scores on its own benchmarks (these are vendor-published, not independently verified) and says Astra is the first model to cross the "Critical" cybersecurity threshold in its Preparedness Framework, triggering extra monitoring and deployment safeguards. API pricing is $10 input / $50 output per million tokens (model ID `gpt-6-astra`), rolling out in phases to ChatGPT Plus/Pro/Business/Enterprise, the API, Azure, and AWS Bedrock — with some reported access delays during rollout.

GPT-5.7 shipped September 3 as OpenAI's broadly-available answer to Claude Opus 4.8. Astra is a separate, much larger release — a full version jump, not a point update — that OpenAI is positioning as a shift in what the model is fundamentally for. The simplest way to describe the difference: GPT-5.7 answers your question well; Astra is built to go do the multi-step task itself.

What actually changed

  • Computer use: Astra can operate inside real computer environments — fill out web forms, update records, browse and research, work inside document editors and spreadsheets, build and QA websites, install and troubleshoot software
  • Agentic coding: rather than one-shot code generation, Astra is built to hold a multi-step objective (check existing architecture, implement changes, add tests, run them, fix failures, update docs) across a single delegated task
  • Handles changing requirements mid-task: OpenAI says Astra can incorporate new instructions partway through a task — a UI change, an added constraint — without losing track of the original objective, a known weak point for earlier agentic models
  • Document generation: produces spreadsheets, presentations, and reports that follow an existing template rather than a generic layout, with judgment calls about what to summarise vs. chart vs. put in an appendix
  • Sites: can generate, host, and share a website, web app, or simple game directly from a prompt, including basic frontend QA on what it built

The benchmark numbers — read with the usual caveat

OpenAI reports strong results on FrontierMath Tier 4, ARC-AGI-3, ExploitBench, and OSWorld 2.0 (a computer-use benchmark, where Astra reportedly scores 72.6% against GPT-5.6 Sol's 65.7%, in less simulated task time). Some of these numbers — a few sit at or near 100% — are unusually high even for a frontier release. These are OpenAI's own reported evaluations. Until independent labs and real-world developer testing confirm similar results, treat them as a vendor benchmark, not proof of real-world performance. We'll update this piece as third-party testing comes in.

Why the cybersecurity angle is the bigger story

OpenAI says Astra is the first model to cross the "Critical" capability threshold in its Preparedness Framework for cybersecurity — meaning that, with the right tools and access, it can reportedly discover previously-unknown vulnerabilities and develop novel exploitation techniques against well-defended systems with less human guidance than before. That is a genuinely different category of capability than "writes better code," and it is why OpenAI says Astra ships with additional monitoring and deployment safeguards rather than the standard rollout process.

A model capable enough to independently develop new exploits is not just a bigger ChatGPT — it changes AI security from an abstract future concern into a concrete deployment consideration today, for OpenAI and for anyone building on top of Astra.

Pricing and access

  • API: $10 per million input tokens / $50 per million output tokens, model ID `gpt-6-astra`
  • Available via ChatGPT Plus, Pro, Business, and Enterprise, plus the API, Microsoft Azure, and AWS Bedrock
  • Rollout is phased, not universal on day one — OpenAI's own release notes describe it as an ongoing rollout
  • Some paid users have reported not receiving access immediately despite the announcement — OpenAI has acknowledged rollout issues

Should you use it?

  • Building agentic workflows (coding agents, computer-use automation, multi-step research): worth testing directly — this is the capability class Astra is specifically built for
  • Everyday chat, writing, or simple Q&A: GPT-5.7 or GPT-5 mini remain the more cost-effective choice — you're not paying for capability you won't use
  • Security-sensitive deployments: read OpenAI's own safety documentation before giving Astra broad computer/tool access, given the cybersecurity capability threshold it triggered
  • Anything requiring today, guaranteed access: check current rollout status for your account tier first, given the reported access delays

Related reading

GPT-5.7 Released — Does It Retake the Lead from Opus 4.8? Claude Opus 4.9 Released — Anthropic's Fast Answer OpenAI Operator vs Anthropic Computer Use What Are Background Coding Agents? AI API Pricing in 2026 — The Honest Comparison

Bottom line

GPT-6 Astra is a genuine step change in what OpenAI is building toward — a model meant to carry out multi-step, tool-using work rather than just answer well — and the cybersecurity threshold it crossed is the part worth paying closest attention to, independent of the benchmark chart. The capability claims are OpenAI's own for now; treat them as a strong opening claim pending independent testing, and expect a messier-than-advertised rollout in the first few weeks.