What's this about?
The numbers, from Anthropic's own table. Terminal-Bench-Science 0.1: 52.6% for Fable 5.1 against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Terminal-Bench 4.0 agentic coding: 55.8%, and 60.9% for Mythos 5.1, against Opus 5's 52.3%. GDPval-AA v2 knowledge work: 1853 versus 1723. OSWorld 2.0 computer use: 77.9% partial, 41.7% strict. Humanity's Last Exam: 60.9% with no tools. Anthropic also notes Fable 5.1 at Low or Medium effort matches or beats Fable 5 at much lower cost.
Why is Terminal-Bench-Science the number to watch?
Because it is the benchmark from last week that was built by working scientists from their own research workflows rather than by model developers, and where the best model then managed 30%. Doubling that in one release is either the most significant result in the announcement or a sign the benchmark is easier to fit than it looked. Anthropic's own footnote helps here: the public leaderboard reports Opus 5 at 30.0% and Fable 5 at 21.4%, while their setup reproduces 29.0% and 24.7% — both within noise, which at least means the harness is not doing the work.
What's the community saying?
- The stated standard error is ±3.5–4.5 points per model, which the 52.6-versus-24.7 gap comfortably exceeds.
- Mythos 5.1 outscoring Fable 5.1 on coding by 5 points is the quiet finding — the difference between them is safeguards, so that gap is the measurable cost of the guardrails.
- The anecdote Anthropic leads with: Millennium reports Fable 5.1 found the cause of a rare crash in its internal systems that no engineer and no other model had explained in several years.