OpenAI released GPT-6 Astra to the general public on 2026-09-04, announcing it as its most intelligent model yet, with reasoning benchmarks it says beat both GPT-5.6 Sol and Claude Fable 5. It is a new model line, not an update to Sol, and it is priced like one: on the API it lists at $10 per million input tokens and $50 per million output tokens, two and a half times Sol and the same as Claude Fable 5.1. We graded it the day it reached Codex.
What the results show
- It is no better at fixing bugs or writing plans than Sol. On the same tasks GPT-6 Astra and GPT-5.6 Sol traded a task here and there and wrote plans of the same quality, so it lands on the same overall grade. It stays clearly behind Claude Fable 5.1.
- It costs more to run. On the $20 ChatGPT Plus plan a run of our suite takes about 15% of the weekly limit with Astra, against 12.6% with Sol. Against Fable it is the cheaper one: at the same monthly price a run takes about 3% of Astra's weekly limit and about 20% of Claude Fable 5.1's.
- It is quicker. The median issue took 1m 56s against Sol's 2m 50s, and a vibe-coding fix 2m 54s against 6m 12s. Planning was a little faster too, at 4m 0s against 4m 46s.
Our take: there is no good use case for GPT-6 Astra right now. Sol is already good and fast enough for everyday work, and for the genuinely hard problems Fable is the model to reach for. This release changes nothing about our ratings. Reports online say Astra is strong at design work and at using a computer on its own, and that may well be true, but our suites do not test either. Astra also works differently from the GPT-5.6 models under the hood, and Codex has not yet been tuned around it, so we expect its grades to improve over the coming weeks as the harness catches up.
Grades are comparable within the same suite version. See the full GPT-6 Astra scorecard or follow the field over time on Trends.