GPT-6 arrives claiming a 10% edge over Fable 5.1. Ten percent of which benchmark is the whole story

GPT-6 has been presented, and according to the brief available to us it improves significantly on previous models including GPT-5.6 Sol and Anthropic's Fable 5.1, with a benchmark advantage described as greater than 10% over Fable 5.1. That is the entirety of the verified material: no named eval suites, no methodology, no independent runs. We're flagging that upfront rather than padding around it.

Because the missing detail is not a detail. On saturated leaderboards — the ones where every frontier model already clusters near the ceiling — a ten-point move is close to noise, or a reflection of how the scoring was framed. On the tests that still discriminate, such as hard software engineering tasks, long-horizon terminal work, or unguided expert security tasks, a 10% jump is a genuine generational step. Same number, opposite meanings, and nothing in the announcement lets us tell them apart.

For context from our own tracking: our Quality index (one third SWE-bench-Pro, one third LMArena, one third Terminal-Bench) had Anthropic's Fable line at the top with 90, ahead of GPT-5.6 Sol at 79. OpenAI has been the challenger at the frontier, so a leapfrog is entirely plausible — labs do overtake each other, and this race has changed leaders before. But a launch-day, first-party margin is the least reliable moment in any model's public life, and we've watched enough Elo scores optimised for voter preference to know the difference between a better model and a better-presented one.

Our reading: this is the evaluation crisis in miniature. With the public thermometers broken, the announcement increasingly substitutes for the proof, and we're asked to accept a percentage in place of a demonstration. What we'll watch over the next weeks: independent reruns on contamination-resistant tests, behaviour on long and messy agentic work rather than clean one-shot problems, and the unglamorous numbers — price per token, rate limits, latency — that determine whether a capability gain reaches anyone. The competitive pressure itself is good news for the long horizon; it's what drives these systems toward the problems that actually matter, starting with biology and medicine. Until someone outside OpenAI measures it, though, "more than 10%" is a marketing figure, not a finding.