Claude Fable 5.1 vs Opus 5: Which One Should You Actually Run?

Anthropic released Claude Fable 5.1 on September 1, 2026. It sits above Opus 5 in the model lineup, costs exactly twice as much per token, and Anthropic's own documentation tells you not to use it by default.

That last part is the whole story, and most of the comparison posts published this week skip it. Anthropic's guidance is to start with Opus 5 and move up to Fable 5.1 only when your evals on Opus 5 at higher effort still come up short. This is not marketing modesty. It is a routing recommendation, and it happens to be correct for most workloads.

Here is what the numbers actually say, including one pricing detail that flips the comparison in a narrow but real set of cases.

The spec sheet

Both models share more than they differ on.

Same context, same output ceiling, same default effort. The differences that matter are price, latency, and how much of a long job each one can carry without falling over.

The cache read inversion, and when it matters

Every line in that pricing table doubles for Fable 5.1 except one. Cache reads run at $0.25 per million tokens against Opus 5's $0.50 — half the price, not double.

This is the only place where the cost comparison genuinely reverses, so it is worth doing the arithmetic instead of hand-waving at it.

For a single call, Fable 5.1 comes out cheaper only when:

cached tokens > (20 × fresh input tokens) + (100 × output tokens)

Run that against a realistic agent loop. If each turn sends 2,000 fresh input tokens and generates 2,000 output tokens, your cached context has to exceed roughly 240,000 tokens before Fable 5.1 is the cheaper call. Push output to 5,000 tokens a turn — normal for code generation — and the threshold jumps past 500,000 tokens.

So the shape that favours Fable 5.1 is specific: an enormous, stable context that gets re-read many times, paired with short outputs. Think a cached codebase or document corpus that an agent queries repeatedly. Retrieval-heavy work with terse answers.

Two things spoil it. Cache writes also double on Fable 5.1 ($12.50 against $6.25 for the 5-minute window), so any workload that keeps invalidating and rewriting its cache pays the penalty twice over. And output tokens carry a 100× weight in that formula, which means verbose generation drowns the cache advantage almost immediately.

For most teams, this is a footnote. If you are running long-horizon agents over a fixed corpus, it is the number that decides your bill.

What the benchmarks show

Anthropic's launch figures put Fable 5.1 well ahead of its predecessor. On Terminal-Bench 4.0 it scores 55.8% against Fable 5's 42.0%, and on Terminal-Bench-Science 0.1 it reaches 52.6%, more than double the earlier model. Those are large jumps by frontier-model standards.

Against Opus 5, the gaps are narrower and uneven. Anthropic reports Fable 5.1 leading on agentic and long-horizon work, with the widest margin on scientific research tasks. On everyday reasoning the difference compresses to a point or two.

Treat all of that as directional. It is vendor-run evaluation, measured with production safeguards active, which Anthropic notes can suppress scores on security-adjacent and biology tasks where the safeguards intervene.

The more interesting data so far is independent. Snorkel AI ran both models against its own frontier coding task set and found Fable 5.1 was not a straight upgrade. Its successful runs were dramatically leaner — the median run used 58% fewer output tokens and finished 36% faster in wall-clock time — but Opus 5 was more robust across the full matched set, with Fable's failures clustering in terminal-heavy and build or dependency tasks.

That result matches the pricing logic rather than contradicting it. Fable 5.1 solves things in fewer tokens when it solves them. It just does not solve strictly more of them.

How to actually route between them

Skip the "which is better" framing. Both are frontier models. The question is which one each workload should hit.

Default to Opus 5 for:

• Bounded tasks with clear acceptance criteria — a patch with a failing test, a scoped refactor, a defined spec
• Anything latency-sensitive, since Opus 5 is documented as moderate against Fable 5.1's slower
• Workloads where you want the option to turn thinking off, which Opus 5 permits at effort high or below and Fable 5.1 does not
• Everyday coding, where the measured gap does not justify a doubled bill

Escalate to Fable 5.1 for:

• Multi-hour agent runs that have to stay coherent across many tool calls and recover from failed steps
• Multistep research where the model needs to hold a thread over a long horizon
• Cases where your Opus 5 evals at high effort still miss the bar, and failure is expensive enough that retries and human review cost more than the token delta
• Large-cached-context workloads that clear the break-even threshold above

A hybrid split works well in practice: Fable 5.1 handles planning and review, Opus 5 does the bulk of execution. You pay the premium on the small fraction of calls where judgment matters most.

Before you commit either way, tune effort. Moving Opus 5 up an effort level is usually cheaper than moving to a model that costs twice as much per token.

Three breaking changes if you are already on Fable 5

Migrating from Fable 5 rather than Opus 5? Three changes will break existing code:

1. Forced tool use returns an error. If your integration pins a specific tool, it needs rewriting.
2. Earlier models cannot read Fable 5.1's thinking blocks. Multi-model pipelines that hand thinking between models will fail.
3. Editing earlier turns invalidates thinking blocks. Any pattern that rewrites conversation history loses them.

Additive changes are more welcome: per-message effort (beta), turn-scoped system messages (beta), readable progress updates between tool calls, the lower cache read price, and content provenance watermarking.

The short answer

Run Opus 5. It is half the price, faster, and close enough on most measured tasks that the gap rarely pays for itself.

Escalate to Fable 5.1 when you have evidence — your own evals, not a benchmark table — that Opus 5 at high effort is failing on work where failure is costly. Or when your context is huge, stable, and re-read often enough that the cache read price does the work for you.

That is Anthropic's own recommendation, and every credible independent test published since launch lands in the same place.

FAQ

Is Claude Fable 5.1 better than Opus 5?

On Anthropic's benchmarks, yes — particularly for agentic, long-horizon, and scientific work. But independent testing found Opus 5 more robust across a full matched coding set, and Anthropic itself recommends Opus 5 as the default. Better on a benchmark table is not the same as better for your workload.

How much more does Fable 5.1 cost?

Exactly double on base rates: $10/$50 per million input/output tokens against Opus 5's $5/$25. Cache reads are the exception, at $0.25 against $0.50.

When is Fable 5.1 actually cheaper than Opus 5?

Only when cached tokens exceed roughly 20× your fresh input tokens plus 100× your output tokens per call. In practice that means a very large stable cached context combined with short outputs.

Do they have the same context window?

Yes. Both offer a 1M token context window and 128K max output. Opus 5 additionally supports up to 300K output tokens through the Batch API beta.

Can I use Fable 5.1 on AWS or Google Cloud?

Yes. It is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The model ID is claude-fable-5-1 everywhere except Bedrock, which uses anthropic.claude-fable-5-1.

What is Claude Mythos 5.1?

The same underlying model as Fable 5.1 with different safeguards, available only to participants in Anthropic's Project Glasswing trusted access programme. It shares Fable 5.1's specifications and pricing.

Continue with AI Tools, the blog archive, AI Tools & Workflows.