Anthropic just shipped Claude Fable 5.1, the first major revision of the Fable line since June. The pitch is simple: same list price as Fable 5, cheaper cache reuse, stronger long-running agents, and fewer of the reliability problems that made the original Fable feel powerful but expensive to keep in the loop.

If you already treat Fable as “the model you bring out when Opus isn’t enough,” 5.1 is the version that makes that choice easier to justify.

Where Fable sits in the Claude stack

Claude’s public lineup is no longer just Haiku / Sonnet / Opus. Fable is the tier above Opus: a Mythos-class model with extra safety wrapping so it can ship to everyone. Mythos 5.1 is the same weights with looser safeguards, and it stays invitation-only through Project Glasswing.

Anthropic’s own guidance is still conservative. Start most work on Claude Opus 5. Move to Fable 5.1 when the job is long-horizon, multi-hour, or when Opus at high effort still misses the bar.

Fable is slower and more expensive. That is the point. It is built for work that used to fall apart after a few hours.

What actually changed from Fable 5

The headline numbers look incremental. The useful changes are operational.

Cheaper reuse. Base input and output prices did not move. Cache reads dropped to $0.25 per million tokens — a quarter of the previous Fable cache-read rate. Anthropic says that cuts typical billed workloads by about 25%, and highly agentic sessions by as much as 45%, because long jobs keep rereading the same prefix.

Better long sessions. Docs and the launch post both emphasize agentic coding over hours, multi-file features, refactors, debugging, review, multistep research, and document / spreadsheet / slide work. The model is also described as more honest when it is stuck and less likely to “pass” a task by disabling a failing test.

Always-on adaptive thinking. There is no non-thinking mode. You steer depth with the effort parameter. Default effort is high. A beta now lets you change effort mid-conversation without blowing the prompt cache.

Tighter thinking rules. Thinking blocks are bound to the model that produced them. Earlier Claude models cannot read Fable 5.1’s blocks. Editing earlier turns can invalidate later thinking. Forced tool use (tool_choice: any or a specific tool) now returns an error. If you have Fable 5 agents in production, those three changes are the migration work.

The capability claims worth paying attention to

Anthropic is selling Fable 5.1 as a coding-and-knowledge-work model first, with research as the teaser for what comes next. A few of the published numbers:

• Terminal-Bench-Science 0.1: 52.6% vs 24.7% for Fable 5
• Terminal-Bench 4.0 (agentic coding): 55.8% for Fable 5.1, 60.9% for Mythos 5.1, vs 42.0% for Fable 5
• Humanity’s Last Exam: 60.9% without tools, 65.0% with tools
• OSWorld 2.0: 77.9% partial / 41.7% strict
• GDPval-AA v2: 1853 vs 1723 for Fable 5 and 1824 for Opus 5

Treat vendor benches as directional. The more interesting part of the launch is the customer anecdotes: a rare crash Millennium’s engineers had not explained, a three-day unattended prototype at MongoDB, kernel work that cut compute cost, and several teams saying they will move traffic off Opus 5 because Fable 5.1 is now close enough on price-per-successful-run. Those are not proofs. They are a consistent story: this model is meant to own a project, not just answer a prompt.

On the research side, Anthropic is more careful. Fable 5.1 can surface vulnerabilities but is not supposed to develop exploits. Some biology and cyber tasks still get routed to Opus. Mythos 5.1 is the variant for vetted life-science and security programs.

Pricing, plans, and the “Covered Model” catch

API pricing:

• Input: $10 / MTok
• Output: $50 / MTok
• 5-minute cache write: $12.50
• 1-hour cache write: $20
• Cache read: $0.25
• Batch: half price on input and output
• US-only inference: 1.1× on input and output

On Claude itself, Fable 5.1 is available to Pro, Max, Team, and Enterprise. On Max and premium Team/Enterprise seats it is part of the plan. On Pro and standard Team seats it is not inside the included limits; you burn extra usage credits. It was not part of the earlier Fable 5 credit promotion.

There is a policy cost as well as a token cost. Fable 5.1 is a Covered Model: extra retention and review rules apply wherever it runs. Default safety monitoring still means data can be kept for up to 30 days on some hosts. Enterprise Frontier Safeguards — customer-cloud storage and a path to zero retention — start rolling out in the fall. If you are in a regulated environment, that is the part of the announcement to read twice.

Who should actually switch

Switch, or at least eval, if you:

• Run multi-hour coding agents against a real repo
• Need the model to keep state, recover, and report progress without babysitting
• Do research synthesis, legal/knowledge work, or slide/spreadsheet production at high quality
• Already pay Fable 5 prices and live on prompt cache

Stay on Opus 5 if you:

• Care more about latency and unit cost than peak reliability
• Do not have long agent loops
• Cannot accept Covered Model retention rules yet
• Have tool-calling code that depends on forced tool use

Developers should update the model ID to claude-fable-5-1 (or anthropic.claude-fable-5-1 on Bedrock), drop forced tool use, keep conversation history append-only if they want thinking blocks to survive, and re-run evals at more than one effort level. Low and medium effort are where Anthropic claims Fable 5.1 can match or beat Fable 5 for less money. High and extra-high are where the quality gap shows.

The short version

Fable 5.1 is not a new species of model. It is Fable 5 with better long-horizon behavior, cheaper cache hits, slightly stricter API rules, and a public-safety wrapper that still leaks some tasks to Opus. That is a less flashy launch than “new generation,” and probably a more useful one.

The test is not whether it wins another bench. It is whether a team can leave it on a hard job overnight and trust the artifact in the morning. Anthropic is betting that this is the first version of Fable that passes that test often enough to become default infrastructure instead of a special occasion.

Released September 1, 2026. Model ID: claude-fable-5-1. Not scheduled for retirement before September 1, 2027.