OpenAI ships GPT-6 Astra, its answer to Claude Fable

OpenAI released GPT-6 Astra, the model it has been trailing for months as the successor to the GPT-5 line and its direct competitor to Claude Fable. The API price lands exactly on Fable 5 and 5.1: $10 per million input tokens and $50 per million output. It is rolling out now to ChatGPT Plus, Pro, Business, and Enterprise, and through the OpenAI API and AWS.

The headline numbers are lopsided. Astra hits 100% on OpenAI's needle benchmark across the 256K to 512K token range, and 100% on ExploitBench, a set of security tasks. On coding it reaches Fable 5's quality for about half the cost. It also posts 99.9% on ARC-AGI-3, though as Simon Willison notes, that figure uses a custom "Provider Adapter" harness that preserves the model's hidden reasoning state between requests and compacts longer conversations, rather than the default setup.

The catch is that raw intelligence is not where Astra wins. On Artificial Analysis's Intelligence Index, Claude Fable 5 still leads it, 66 to 61. So the picture is a model tuned hard for security, long-context retrieval, and cost-efficient coding, priced to sit right next to the model it is chasing.

Why it matters

If you already build on Fable or GPT-5.6, Astra changes the math for two workloads specifically: long-context retrieval and coding, where it claims Fable-level output at half the price. Test those two before switching, and don't assume the general reasoning gap closed just because the benchmark charts look dramatic.