GPT-6 Astra arrives with more autonomy and higher rates

OpenAI released GPT-6 Astra on September 3, starting a phased rollout focused on work that takes multiple steps on a computer.¹ ⁵ The company reports 57.9% on Terminal-Bench 4.0, against 37.3% for GPT-5.6 Sol. Those are the highest scores achieved at any effort level in its published tests; they don’t isolate Low against High.¹

Astra Low can outperform Sol High

OpenAI’s usage guide says Astra at Low can outperform Sol at High. It suggests Low or Medium as a starting point for people already getting good results from Sol High. Effort is a separate setting from model choice, and High can consume more allowance without improving an answer.²

That gives users a concrete comparison to run before increasing effort. Keep the request, files and acceptance criteria the same. A bug fix that passes the existing tests offers a clearer comparison than judging two answers by their length.

Token rates and task costs can move differently

The API model page lists $10 per million input tokens and $50 per million output tokens, with separate cache rates. Requests above 272,000 input tokens incur twice the input and cache rates and 1.5 times the output rate for the entire request. The context window holds 1.05 million tokens, with a 128,000-token output limit.³

OpenAI’s model guide says Astra uses fewer output tokens in some evaluations, lowering estimated API cost per task despite higher unit prices.⁴ Whether that saves money depends on the finished work, the retries and the volume processed. A short answer that forces the task to be repeated still carries a cost.

For subscribers, Work and Codex share an allowance. OpenAI estimates 5–45 local Astra messages per five-hour window on Plus, versus 10–100 for Sol. These ranges aren’t fixed message caps.²

Ongoing work needs ongoing oversight

The API adds asynchronous tool calls and instructions sent while work is underway. An agent can advance independent parts of a task while a tool runs, then incorporate a correction without discarding completed work. The application still executes the tools and manages pending operations.⁴

OpenAI places Astra at the Critical cybersecurity capability level under its own framework and has added monitoring to external tool-using runs. It also reports that written reasoning became harder to monitor in adversarial tests, even as its broader evaluations show better adherence to restrictions than Sol.⁵ Greater ability to act makes the actions taken part of the result that needs evaluating.

Sources

1. GPT-6 Astra: A new generation of intelligence · OpenAI · September 2026 launch.
2. Managing usage with GPT-6 Astra in Work and Codex · OpenAI Help Center · accessed September 9, 2026.

Show 3 more sourcesHide sources

3. GPT-6 Astra Model · OpenAI API · accessed September 9, 2026.
4. Model guidance — Using GPT-6 Astra · OpenAI API · accessed September 9, 2026.
5. Safety overview: GPT-6 Astra · OpenAI · September 3, 2026.