GPT-6 Astra Review

Astra aims at one thing: taking a messy multi step job, driving software to do it, and finishing without a person watching. On that it is the strongest release of the year. It clears desktop task suites well above the model it replaces and does it in roughly half the time per task, terminal work and browsing improve in the same direction, and the practical effect is that a category of work which used to need supervision now needs a review at the end.

OpenAI also gates it, holding the cyber capabilities behind a trusted access program after its own evaluation put the model at the critical threshold for finding and chaining unknown vulnerabilities, which is the first time a lab has slowed its own launch on that basis. The cautions are real. Every headline result so far comes from OpenAI's own harness with no outside replication, and on the independent aggregate index and on the hardest open exam it sits behind Anthropic's current top model rather than ahead of it.

It costs about twice what the previous flagship does per token, and it arrived as an approved access preview a day before anyone else could buy it, which is where ease of use loses its tier. For agentic work it is the first call. For everything else the price argues otherwise.