The launch-day number everyone quoted was SWE-Bench Verified. The more telling one is SWE-Bench Pro — the harder, contamination-resistant variant — where Fable 5 posts 80.3%. The rest of the field trails by double digits: Opus 4.8 at 69.2%, GPT-5.5 at 58.6%, Gemini 3.1 Pro at 54.2%.
The pattern to notice isn't the lead; it's that the lead widens as tasks get harder. On FrontierCode's Diamond split — the hardest problems in Cognition's frontier coding eval — Fable 5 scores 29.3% against Opus 4.8's 13.4% and GPT-5.5's 5.7%. It more than doubles the second-place model precisely where the others fall apart. Cursor's Michael Truell calls it "the state of the art model on CursorBench," crediting its long-horizon reasoning.
That shape — modest edge on easy work, commanding edge on long-horizon work — matches the field reports. Fable 5's advantage shows up most in agentic settings: fewer wasted tool calls, more persistence, reading its own error output and revising instead of stopping to ask. The headline case remains Stripe's, which Anthropic says used Fable 5 to migrate a 50-million-line Ruby codebase in a day — work estimated at two-plus months by hand.
The caveat worth keeping: several third-party evaluators note that on the longest agentic tasks, runs hit the harness timeout before finishing — so some results read as an experience signal rather than a clean leaderboard score. The evidence wall, with every claim and its receipt, is on /benchmarks/.