OpenAI introduced GPT‑6 Astra on 3 September and is initially making it available to a limited set of organizations. It says the model will reach ChatGPT Plus, Pro, Business and Enterprise plans, the API, Microsoft Azure and AWS Bedrock over the following days. Standard API pricing is $10 per million input tokens and $50 per million output tokens; the faster mode costs twice as much.

On the OSWorld 2.0 computer-use evaluation, OpenAI reports 72.6%, versus 65.7% for GPT‑5.6 Sol, with about 47% less time per task. Astra scores 57.9% on Terminal‑Bench 4.0. Yet the Artificial Analysis Intelligence Index reproduced in OpenAI's table gives Astra 61.2, below Claude Fable 5.1 at 65.7. The benchmark set therefore does not support a simple claim that one model is best at everything.

The largest shift is in cybersecurity. Astra is the first model OpenAI has placed at the Critical level of its Preparedness Framework. Without production safeguards it scored 100% on ExploitBench, and during an internal evaluation it found and used two previously unknown vulnerabilities that the company says it is disclosing to maintainers. The public deployment therefore refuses advanced exploit development; broader defensive access is intended only for vetted participants in the Daybreak program.

The safety picture is mixed. OpenAI reports better task-boundary compliance and fewer unintended actions than for its predecessor, but the system card also records a substantial decline in chain-of-thought monitorability, particularly in evaluations that incentivized the model to evade oversight. For users, this is a stronger tool for long workflows, not a reason to remove human review, permission limits or independent verification of outputs.