GPT-6 launches claiming a 10% edge over Fable 5.1 — the number that matters is which benchmark

OpenAI has presented GPT-6, positioned as a clear step up from both its own GPT-5.6 Sol and Anthropic's Fable 5.1. According to the launch material, the model shows a benchmark advantage of more than 10% over Fable 5.1. That is the extent of what has been made public so far: a launch, a comparison, and a percentage.

Context matters here, because that percentage would represent a genuine reversal. On our own Zendoric Quality index — a composite of SWE-bench-Pro, LMArena and Terminal-Bench — Anthropic's Fable 5 sat at 90 against GPT-5.6 Sol's 79, an eleven-point gap that was the most stable feature of the frontier leaderboard. A GPT-6 that beats Fable 5.1 by double digits does not merely close that gap; it inverts it. Lead changes of that size at the top of the frontier have been rare, which is precisely why the claim deserves scrutiny rather than applause.

And scrutiny means asking the question the announcement does not answer: which benchmarks. "More than 10%" is a meaningless figure until you know whether it comes from tests that still discriminate between strong models or from tests that everyone already aces. We have made this point repeatedly about cybersecurity evaluations — Cybench is effectively saturated at near-100% pass rates, so a win there tells you nothing — and the same logic applies to coding and reasoning suites. Averaged composites reward a model that is broadly competent; expert unguided tasks and real software-engineering benchmarks reward a model that is actually better. A lead built on the former is marketing. A lead built on the latter is capability.

Our read: treat this as a credible signal and an unverified claim at the same time, and resist the reflex to rewrite the leaderboard on a press release. If GPT-6's margin survives independent testing on the hard tasks, the practical consequence is not bragging rights but pressure — on Anthropic to answer, on Google to stop trailing on the difficult evaluations, and on the Chinese open-weight labs (GLM, Qwen, DeepSeek, Kimi) whose strategy has been to shadow the closed frontier at a fraction of the cost. Competition at this altitude is the mechanism that turns frontier capability into cheap, widely available capability, and cheap capability is what eventually reaches a diagnostic clinic or a drug-discovery pipeline rather than staying inside a demo. That is the long game worth caring about.

The honest caveat is that we are commenting on a thin factual base: a launch and one relative figure, with no independent evaluation and no disclosed test set. We will revisit GPT-6's position on our indices once there are numbers we can verify rather than numbers we have been handed.