September 24, 2026
GPT-6 Astra speeds up AI agent work: 72,6% in OSWorld
GPT-6 Astra completed OSWorld 2.0 tasks in roughly 40 minutes instead of 75 for GPT-5.6 Sol: in OpenAI’s September 3 simulation, the model scored 72,6% versus 65,7%. On September 24, Astra is available in ChatGPT Work and Codex for Plus, while GPT-6 Pro based on it is included in Pro for $100 and $200, Business, and Enterprise.
Peter Steinberger 🦞
@steipete
Astra blows everything else away.
· 11.2K views
Previously, GPT-5.6 Sol scored 37,3% on Terminal-Bench 4.0. OpenAI reports 57,9% for Astra, versus 55,8% for Claude Fable 5.1, and estimates API tasks to be 9% cheaper than Sol and 63% cheaper than Fable.
Long tasks. Through the API, Astra is called as `gpt-6-astra` only through the Responses API with tool calling. Its context holds 1 050 000 tokens, roughly an average novel in full, and responses can be up to 128 000 tokens long. The model has five reasoning levels: from `low` to `max`.
Price and launch. One million input tokens cost $10, output tokens cost $50, and cached input costs $1. After 272 000 input tokens, OpenAI doubles the price of input and cache, while output becomes 1,5 times more expensive. Astra in Codex requires CLI 0.153.0; when migrating, `reasoning.effort: "none"` changes to `low`, and running tests and fixing failures should be explicitly included in the task.
If Astra has not appeared in ChatGPT Desktop, OpenAI recommends updating the app through Menu → Check for Updates.
Source
