GPT-6 Astra ties Claude Fable 5.1 in Artificial Analysis' intelligence and coding-agent indices, according to the evaluator's September 9 results. In the Intelligence Index, Astra at maximum reasoning effort scores 53, matching Fable 5.1 at maximum effort with fallback. The reported cost is $3.26 per task for Astra versus $7.63 for Fable.
Lower token use helps explain the difference. Astra produces about 27,000 output tokens per intelligence task, compared with Fable's 78,000 at the tested settings. These are evaluation costs, not the price of a typical chatbot conversation.
In the Coding Agent Index, Astra running in Codex and Fable running in Claude Code both score 62. Astra costs $7.09 per task, roughly 40% less than Fable in that comparison. Different agent environments are part of those results; this is not a comparison of bare models under one identical harness.
Astra also has regressions against its predecessor. Its DeepSWE score falls from GPT-5.6 Sol's 72% to 68%, and Artificial Analysis reports a roughly 45-point Elo decline on GDPval-AA v2. The combined index therefore does not establish an improvement on every kind of coding or professional task.