Claude Fable 5.1 vs Opus 5 vs Sonnet 5: Which Model Should You Actually Use?

Four models make up the current Claude lineup. All figures below are from Anthropic's own model comparison table, current as of this article's publish date.

Two things worth noticing before any benchmark discussion. First, Fable 5.1 costs exactly 2x Opus 5 on both input and output tokens, that's the price of admission for whatever the top tier buys you. Second, all three non-Haiku models share the same 1M-token context window and 128K max output, the differentiation between them isn't capacity, it's reasoning depth and, correspondingly, cost.

Anthropic's own "Choosing a model" guidance states the default plainly: "start with Claude Opus 5 for most workloads. Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short." That's not marketing copy, it's the vendor recommending you not use its own most expensive model by default.

Fable 5.1 shipped September 1, 2026, as a point release over the original Fable 5. Pricing is unchanged ($10/$50 per MTok), the changes are in capability, cache economics, and safeguard tuning.

Benchmark deltas vs Fable 5

The pattern is consistent and worth sitting with: the two benchmarks that measure long-horizon, multi-step agentic work (Terminal-Bench-Science, Terminal-Bench 4.0) show large gains, while the two that measure more general reasoning and single-session coding assistance (CursorBench, HLE) show gains of a few points. Fable 5.1's improvements are concentrated in exactly the kind of work its name implies: long-running, autonomous, multi-step tasks, not everyday coding assistance.

Cache pricing cut

Cache reads dropped to $0.25 per million tokens, a 75% reduction (prior cache-read pricing at the $10 input tier worked out to roughly $1.00 per MTok, the standard 10%-of-input-price rate). Anthropic states this yields "around 25%" cost reduction for typical workloads and "up to approximately 45%" for highly agentic work that reuses context heavily across many tool calls.

Safeguard tuning

Biology-related safeguards now fire roughly 85% less often on benign biology and medical questions, and cybersecurity safeguards fire about 60% fewer interventions per session, changes Anthropic attributes to more precise classifiers rather than loosened policy, paired with expanded allowance for defensive security work.

Availability

Fable 5.1 is available on the Claude API, Claude Code, Claude Enterprise, and Claude Platform, plus AWS, Google Cloud, and Microsoft Azure. Mythos 5.1, the higher-safeguard sibling built for cybersecurity and life-sciences work, remains restricted to vetted organizations under Anthropic's Project Glasswing program.

Put the pricing table and the benchmark table side by side and a practical rule falls out. Fable 5.1 earns its 2x price premium over Opus 5 specifically on tasks that resemble Terminal-Bench-Science and Terminal-Bench 4.0: long-horizon, multi-step, largely autonomous work where the model operates for many turns with limited human checkpointing, exactly the profile Anthropic's own docs describe as "demanding reasoning and long-horizon agentic work."

For everything that looks more like CursorBench or HLE, single-session coding help, focused reasoning tasks, question answering, the measured gap is a few points, not enough to justify doubling the token bill by default. This is also consistent with a detail buried in the API surface: Fable 5.1 has thinking permanently on, no way to disable it, and a mandatory 30-day data retention requirement with no zero-data-retention option. Anthropic built it for sustained, deeply reasoned, retained-context work. Using it for quick single-turn requests pays the always-on thinking tax without the long-horizon payoff that justifies it.

A practical framework, in order of what to check first:

1. Start with Opus 5 at effort `high`. This is Anthropic's own stated default, and at half Fable 5.1's price with the same context window, it's the right baseline for the large majority of coding and reasoning work.
2. Move to Fable 5.1 only after Opus 5 falls short on your own evals, not on a leaderboard. Anthropic's own guidance frames this as the actual trigger condition, "when your evals on Claude Opus 5 at higher effort still fall short," not a benchmark score threshold.
3. Reach for Fable 5.1 directly for genuinely long-horizon agentic tasks, multi-step research, large autonomous coding sessions, work that resembles Terminal-Bench-Science or Terminal-Bench 4.0 more than a single Q&A turn.
4. Use Sonnet 5 where latency matters more than the last few points of capability. Same 1M context window and 128K max output as the top two tiers, at $2/$10, with faster comparative latency.
5. Use Haiku 4.5 for high-volume, latency-sensitive, or simple-task routes where 200K context is enough and $1/$5 pricing matters at scale.

A handful of API-level differences on Fable 5.1 are easy to miss when migrating code written against Opus 5 or an older model.

Thinking cannot be disabled. Fable 5.1's thinking is adaptive and always on, omit the parameter or pass {type: "adaptive"}; explicitly sending {type: "disabled"} returns a 400 error. Code carried forward from a model where disabling thinking was a valid cost-saving move will break.

No assistant message prefill. Fable 5.1, along with the rest of the current-generation lineup, rejects prefilled assistant turns with a 400. Response-format control has to go through structured outputs (output_config.format) or system-prompt instructions instead.

30-day data retention is mandatory. Fable 5.1 is not available under zero data retention, a request from an organization whose retention configuration doesn't meet that minimum returns a 400 invalid_request_error. Confirm this is acceptable for your org before switching a production route to Fable 5.1.

`refusal` is a distinct stop reason to handle. Safety classifiers can decline a request with HTTP 200 and stop_reason: "refusal" rather than an error, code that only checks for exceptions will silently treat a refusal as a normal, empty-ish response instead of surfacing it.

None of these are reasons to avoid Fable 5.1 for the workloads it's built for. They're reasons not to swap the model string in existing Opus 5 or Sonnet 5 code and assume nothing else needs to change.