GPT-6 Astra vs Fable 5.1: a benchmark-by-benchmark comparison
A long post compares the two models on computer use, coding, cybersecurity, long context, hallucination rate, and price per million tokens.
ran GPT-6 Astra and Fable 5.1 side by side across the benchmarks here's what actually separates them GPT-6 Astra: - computer use, runs the table, OSWorld at 72.6%, finishes tasks 47% less time than its own predecessor - coding, Terminal Bench 4.0 jumped from 37.3 to 57.7 gen over gen, DeepSeek's long-agentic benchmark hit 74.1% - cybersecurity, exploit bench went from 5.5% to 39%, a full step change, this is also why rollout is throttled - long context, 100% on the eight-needle test up to 512K tokens, still 96.3% at a full million - hallucination rate dropped from 9.4% to 2% - $10/million input tokens, $50/million output, and token-efficient enough that a ~55% Terminal Bench score costs $7.20 instead of Fable's ~$20 Fable 5.1: - wins the artificial analysis intelligence index outright, 65.7 vs Astra's 61.2 - scores competitively on Frontier Code 1.1, actually ahead of Astra at the high end in one config - costs more per accuracy point across nearly every benchmark where it's directly compared - rollout is already broad, no staggered access, no banked-reset compensation scheme the real fork isn't raw intelligence, it's cost-to-accuracy and availability Astra: better token economics, faster, safer against exploit misuse, but still staggering out to users and priced the same per token despite scoring better Fable: wins on one specific reasoning index, available now, no waiting for anyone running high-volume agentic workflows, coding or computer-use heavy, Astra's cost-per-point advantage is the actual decision driver for anyone who needs the model today with no rollout lag, Fable is already sitting there
This post quotes another post: x.com/scaling01/status/2095577644431515887