What changed with Fable 5.1, and what I do differently because of it
Anthropic launched Claude Fable 5.1 (alongside Mythos 5.1) on September 1, 2026. Bloomberg's framing: better at coding and science, cheaper to run. The staged rollout had been visible since late July, in Bedrock backend changes and gray-scale tests on Claude Web and Claude Code.
Below is what actually changed, and next to each one, what it changes about how I use the model.
The list
1. Built for long-horizon agent work, not single-shot answers. It plans multi-step tasks, recovers from a failed step without a human, and checks its own output with vision against the stated design goal before calling a task done. Use it differently: stop babysitting one prompt at a time. Hand it the whole task with an explicit success criterion, and let it run unattended across tools (Slack, browser, terminal) instead of chopping the job into supervised chunks.
2. Agentic coding jumped. Terminal-Bench 4.0: 55.8% vs Fable 5's 42.0%, and above Opus 5's 52.3%. CursorBench 3.2.0: 73.4% vs 70.5%. Use it differently: Fable is now my default for the agentic loop, not the model I escalate away from. Opus stops being the automatic "hard problem" fallback.
3. Agentic science research more than doubled. Terminal-Bench-Science 0.1: 52.6% vs 24.7%. Opus 5 sits at 29.0%. Use it differently: research-shaped work that used to be a human-in-the-loop literature crawl is now worth handing over as one delegated task.
4. Business-workflow automation roughly doubled. AutomationBench: 31.4% vs 17.1%. OSWorld 2.0 (computer use): 77.9% vs 72.9%. Use it differently: multi-app workflows that previously failed halfway are worth retrying. Note the absolute number is still 31%, so the retry is worth it, unattended production trust is not.
5. Knowledge work and reasoning up. GDPval-AA v2: 1853 Elo vs 1723. Humanity's Last Exam: 60.9% no tools (vs 57.8%), 65.0% with tools. Use it differently: fewer reflexive escalations to a bigger model for reasoning-heavy one-offs.
6. Fewer tokens for the same answer. At Low and Medium effort it matches or beats Fable 5's accuracy on less. Anthropic's internal finance benchmark: same accuracy, 20% fewer tokens. A hard browser-agent benchmark: 82% of tasks done in ~10 minutes each, against 74% for Opus 5 and 57% for Fable 5, on fewer tokens than either. One customer reported roughly 2x the speed of Opus 5 at about half the tokens. Use it differently: default to Low or Medium effort and only raise it when a task actually fails. High effort is now the exception, not the safe choice.
7. Cache reads dropped 75%. $0.25 per million. Input and output list prices are unchanged at $10 and $50 per million. Use it differently: this is the biggest practical change. Structure prompts so the large stable part (the codebase, the spec, the tool definitions) sits at the front and gets cached, and only the small variable part changes per call. Anthropic's estimate is ~25% off typical workloads and up to ~45% off highly agentic ones. Long agent loops and full-codebase analysis are where it lands hardest.
8. US-only inference is available at 1.1x standard pricing. Use it differently: only relevant if you have a data-residency requirement. Otherwise it is a 10% tax for nothing.
The benchmark table
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% |
| Terminal-Bench-Science 0.1 (agentic research) | 52.6% | 24.7% | 29.0% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | 70.5% | 70.0% |
| GDPval-AA v2 (knowledge work, Elo) | 1853 | 1723 | 1824 |
| OSWorld 2.0 (computer use, partial credit) | 77.9% | 72.9% | 75.4% |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% |
| Humanity's Last Exam (with tools) | 65.0% | 63.8% | 63.6% |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% |
Fable 5.1 beats Opus 5 on most of these, despite Opus 5 having been the top-ranked model on several leaderboards. Worth noting the suite itself moved: Fable 5 set records at its June launch with 95.0% on SWE-bench Verified and 80.3% on SWE-bench Pro, the first model past 80 on Pro. The 5.1 materials lead with newer agentic and science tests instead of repeating those, which is Anthropic signalling where they think the frontier moved.
The one thing that does not change
Every number above is Anthropic's. No independent reproduction yet.
Prior system-card analysis of the Fable 5 generation found the most flattering published results leaned on internal harnesses, in-house grading, and token budgets you cannot reproduce through the public API. That is a whole-industry pattern, not an Anthropic one, but it is the reason I treat launch benchmarks as directional and nothing more until SWE-bench.com, LLM Stats, or Vals AI publish verified scores.
And the token-efficiency claim is the one I would test on my own workload before believing. I ran a cost-router benchmark that killed its own feature: the cheap model cost 2-7x more per task despite listing at half the price, because per-token price is not per-task cost. Same trap applies in reverse here. The way to know whether 5.1 is cheaper for you is to run your own paired tasks, not to read the launch page.