Claude Fable 5: Anthropic's Most Capable Model, Explained

Claude Fable 5 is Anthropic's most capable widely released model, built for the most demanding reasoning and long-horizon agentic work. It is not a general-purpose upgrade path — it is a deliberate top tier, priced and engineered for problems that smaller models cannot finish.

What is actually different

Fable 5 changes several assumptions that hold on every other Claude model.

Thinking is always on. There is no way to disable it. Where earlier models accepted a parameter to turn reasoning off, Fable 5 rejects that outright — reasoning is a permanent property of the model, and depth is controlled through an effort setting spanning low through max rather than a token budget.

The raw chain of thought is never returned. You can request a readable summary of the model's reasoning, but the unfiltered reasoning trace is not exposed under any setting. Reasoning still happens and is still billed identically; what changes is only what you can see. For regulated environments this is worth understanding early, because "show me exactly how the model reached this conclusion" has a different answer here than it does elsewhere.

It requires 30-day data retention. Fable 5 is not available to organisations configured for zero data retention — those requests fail outright, regardless of how well-formed they are. For GCC enterprises operating under strict data-handling mandates, this is the single most important qualifying question, and it is a procurement question rather than an engineering one.

It can decline. Fable 5 runs safety classifiers targeting research biology and most cybersecurity content, and a declined request returns a successful response carrying a refusal signal rather than an error. Benign adjacent work — security tooling, life-sciences research — occasionally triggers a false positive. Anthropic's answer is a fallback mechanism that transparently re-runs the request on another model, which is worth configuring from day one rather than adding after the first incident.

The performance picture

On the Artificial Analysis Intelligence Index, Claude Fable 5 at max effort scores 60. That put it at the frontier at launch, ahead of GPT-5.6 Sol at 59 and Claude Opus 4.8 at 56.

Its strongest showing is in knowledge work. On AA-Briefcase — a benchmark built from realistic professional tasks spanning thousands of input files, requiring deliverables like research reports, presentations, and spreadsheets — Fable 5 recorded an Elo of 1574 and a Rubric Score of 56%, well ahead of GPT-5.6 Sol at 42%. Where Sol produces the better-looking deck, Fable 5 produces the more correct one.

It also retains an advantage in factual knowledge. On AA-Omniscience, which measures both accuracy and hallucination rate, Fable 5 outperforms the smaller Opus-class models — a direct consequence of its size class.

The cost is real: $10 per million input tokens and $50 per million output, roughly double the Opus tier, at $22.30 per AA-Briefcase task. It offers a 1 million token context window — both the default and the maximum — with up to 128,000 output tokens per request.

Where it fits, and where it does not

Fable 5 earns its price on a narrow set of problems: long autonomous runs measured in hours rather than minutes, first-shot implementations of well-specified systems, end-to-end enterprise deliverables, and multi-agent work where a coordinator delegates to and sustains communication with long-running sub-agents.

It is over-specified for routine classification, summarisation, and extraction. Teams that default everything to the top tier will pay several times over for work a mid-tier model completes just as well.

There is a counterintuitive point worth flagging for anyone migrating. Prompts and skills written for earlier models are often too prescriptive for Fable 5, and enumerating steps measurably reduces output quality. The model performs better when told the goal and the constraints and left to determine the method. If you are porting a mature prompt library, A/B test with the step-by-step scaffolding removed before assuming it still helps.

Two operational realities follow from its capability. Individual requests on hard tasks can run for many minutes at higher effort — a fifteen-minute single request is normal — so timeouts, streaming, and progress indicators need planning before rollout, not after. And the model performs materially better when given somewhere to write down what it learns; even a plain markdown file it can consult in later sessions produces a measurable improvement.

The strategic read for GCC enterprises

The arrival of a distinct top tier, priced well above the models beneath it, marks the end of the period where "use the best available model" was sound default advice. That approach now carries a real and avoidable cost.

The organisations getting the most from Fable 5 in early access shared one trait: they pointed it at their hardest unsolved problems first, rather than migrating existing workloads onto it. That is the right instinct. Fable 5 is not a faster way to do what you already do — it is a way to attempt work that was previously out of reach.

Everything else should run on a cheaper tier, with Fable 5 reserved for the problems that justify it.