OpenAI’s GPT-6 Astra raises the ceiling for tool-using AI, but it also raises the per-token price and the operational stakes. The model is rolling out first to a limited set of organizations, with broader ChatGPT and API access promised over the coming days.
What changed
Astra combines a 1.05-million-token context window with computer use, browsing, coding, research, document creation, and the full Responses API tool set. OpenAI says it substantially improves long-running agent work and can preserve searchable notes across Codex context windows instead of depending only on repeated compaction. The company also classifies Astra at the “Critical” cybersecurity capability level under its Preparedness Framework—the first broadly deployed OpenAI model to reach that threshold.
The practical consequence
For difficult workflows, teams can assign longer, more tool-heavy jobs to one model and expect stronger computer-use and software-engineering performance. That does not make Astra the automatic choice for routine traffic. Standard API pricing is $10 per million input tokens and $50 per million output tokens, 2.5 times GPT-5.6 Sol’s listed per-token rates. OpenAI says lower token use can make some evaluated tasks cheaper overall, but real cost will depend on prompts, reasoning effort, tool calls, and retry behavior.
Operational and safety limits
Access is staged rather than generally available on day one, and Enterprise administrators must enable Astra because it starts off by default. Fine-tuning, audio, and video are not supported. Fast processing costs twice the Standard rate, and prompts above 272,000 input tokens trigger higher rates for the full request. The stronger cyber capability also brings stricter monitoring: legitimate work may be slowed or stopped, and an API task can terminate instead of asking a person to approve continuation.
Benchmark status
OpenAI reports large gains on computer-use, coding, science, long-context, and cybersecurity evaluations, including 72.6% on an offline OSWorld 2.0 subset and 100% on ExploitBench. These are launch-day results selected and reported by the vendor; DrComps did not independently reproduce them. Several comparisons use internal evaluations or particular harness settings, so they should not be read as universal workload performance.