Three months after Fable 5's bumpy start (launch, forced shutdown, comeback), Anthropic quietly shipped a version update on September 1, 2026 — and this one actually has teeth. On one of the more important coding benchmarks, Fable 5.1 doubles its predecessor's score. No drama this time, just numbers.
What actually changed
On Terminal-Bench 4.0 (measures how well an AI agent handles command-line tasks on its own), Fable 5.1 scores 55.8%. Fable 5 managed 42.0%, and Claude Opus 5 sat at 52.3%. That means the "smaller" Fable model now beats Anthropic's own flagship Opus — on paper, at least.
The jump is even bigger on Terminal-Bench-Science (tests whether an agent can carry out scientific tasks from the command line): 52.6%, against Fable 5's 24.7%. More than double, in a single version bump.
Alongside Fable 5.1 comes a restricted twin again: Mythos 5.1 — same model, fewer safety guardrails, available only to a narrow, vetted group of cybersecurity and life-sciences researchers. For everyone else, Fable is the one that matters.
Same price, cheaper caching
Base pricing stays put: $10 per million input tokens, $50 per million output tokens — identical to Fable 5. What changes is cached (reused) requests: they now cost 75% less, down to $0.25 per million tokens.
Sounds like a footnote, but it's the real lever for repetitive workloads — think an AI agent that keeps re-reading the same code context. Anthropic estimates roughly 25% savings for typical use and up to 45% for heavily agentic workflows. Run Fable at low or medium "effort" and, according to Anthropic, you get results similar to or better than Fable 5 — for noticeably less money.
Free? Same answer as last time: no
If you were hoping for a free tier: nope, same story as usual in this price bracket. Free accounts on claude.ai don't get Fable 5.1 at all — it stays behind a paid plan. On a Pro plan, it's not included in your regular usage limit; it costs extra usage credits. Only Max subscribers can burn through up to 50% of their weekly limit on Fable models at no extra cost — beyond that, credits kick in too.
This isn't Anthropic being stingy specifically — pretty much every provider keeps its best model behind a paywall, and GPT-6 Astra was no different. If you only want to try it occasionally and don't have a paid plan, you're stuck with whatever your free tier already offers.
Who actually benefits
The clearest winners are developers working through the API with lots of repeated context — that's where the caching discount pays off the most. Max subscribers also feel the difference immediately, at no extra cost. Anyone on a free account, or just chatting occasionally on claude.ai, notices exactly nothing: no new access, no new features in their own plan.
Bottom line: Fable 5.1 is a solid technical update backed by real numbers, not just marketing. It's just that the people who don't already pay still don't get to enjoy it.