Anthropic has rolled out Claude Fable 5.1, model ID claude-fable-5-1, available simultaneously through the Claude API, Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.1, identical in specs and pricing, launched alongside it but is invite-only for now.
The foundations carry over from the previous generation: a flat price across the full 1-million-token context window, a maximum output of 128,000 tokens, adaptive thinking on by default, and a knowledge cutoff Anthropic considers reliable through June 2026.
Only One Cell on the Price Sheet Moved
Input is $10 and output is $50 per million tokens, unchanged from the previous generation, and cache-write pricing hasn't moved either. The one thing that changed is cache-read: Fable 5.1 bills it at 0.025 times the base input price — $0.25 per million tokens — versus 0.1 times for other Claude models.
The discount only pays off for one pattern: reading the same prefix over and over. Rough math: a 200,000-token system prompt plus codebase prefix, re-read 100 times over a multi-hour agent session, adds up to 20 million tokens of cache reads — $20 under the old pricing, $5 under the new. Cache-write pricing and the 512-token minimum cacheable length are both untouched; the entire savings comes from the act of reading back.
Long-running agent sessions are exactly where that re-reading is densest. Aiming the discount at this one cell makes the intent hard to miss.
Three Changes That Will Throw Errors Immediately
Migrating from the previous generation by swapping only the model ID, without touching the surrounding code, will break in at least three places.
Forced tool calling is gone. Setting tool_choice to any or to a specific tool now returns a 400 outright. Anthropic's explanation: this model keeps thinking on by default, and forcing a tool call would skip that thinking step — the model would end up stuffing its reasoning into the tool parameters instead, degrading their quality. For guaranteed-valid JSON, use strict mode or structured output instead.
Thinking blocks now carry model identity. Every thinking block records which model produced it, and that record only travels one way: Fable 5.1 can read thinking blocks from earlier models, but earlier models can't read its. Upgrade mid-session and the reasoning chain survives; downgrade and those turns' reasoning is gone. When routing or fallback logic switches models mid-conversation, the API quietly drops the blocks it can't read — uncharged, and by default without telling you.
Editing history invalidates every thinking block that follows it. Touch anything that comes before a thinking block — the system prompt, the tools array, an earlier message — and the next request returns a 400. This is enforced for any account created on or after August 31. The most common way to trip it: injecting a temporary per-turn reminder into history and deleting it the next turn, or rebuilding the system prompt between requests. Anyone using an official client doesn't need to worry about this; anyone assembling their own message arrays does.
What's Saved on Input Can Come Back on Output
More worth watching than the breaking changes are the default behavior shifts that happen without touching any code. Anthropic lists seven of them, and three go straight to the bill.
First, parallel tool calling is less reliable: where the previous generation would fire off several calls in one turn, 5.1 may send just one at a time — the documentation says answer quality doesn't suffer, but the extra turns cost tokens and round trips. Second, it leans toward rewriting whole files rather than editing pieces; the result is usually the same, but it burns more output tokens getting there. Third, at low effort settings it calls search less often, leaning more on memorized knowledge instead.
Set alongside the price sheet, what 5.1 is really doing looks less like a price cut and more like moving cost from the input side to the turn side. Cache reads are three-quarters cheaper, but choppier tool calls and full-file rewrites eat into the $50-per-million output price — five times the input price. Whether your bill goes up or down depends on which matters more in your specific agent loop: repeatedly re-reading context, or repeatedly re-emitting content. Re-running your evals and reconciling the bill once after migrating isn't a step you can skip.
Two of the four new beta features — changing effort level mid-session, and system messages that apply for a single turn only — exist specifically to work around the pitfalls above. There's also a new default-on provenance feature: generated text carries a statistical watermark across every platform, and retrieved images and video carry C2PA credentials.
The Widest Benchmark Gap Is in Research Agents
In Anthropic's own head-to-head comparison, the gains are far from uniform:
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 (agentic scientific research) | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 (agentic coding) | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (knowledge work, score) | 1853 | 1723 | 1824 | 1711 |
| OSWorld 2.0 (computer use, partial / strict) | 77.9% / 41.7% | 72.9% / 36.1% | 75.4% / 39.6% | — |
| Humanity's Last Exam (no tools / with tools) | 60.9% / 65.0% | 57.8% / 63.8% | 56.6% / 63.6% | — |
| AutomationBench (business workflows) | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 (agentic coding) | 73.4% | 70.5% | 70.0% | 67.2% |
The widest gap is in research agents: 52.6% versus the previous generation's 24.7%, more than double. That's also the one place the previous generation clearly trailed Opus 5 (24.7% vs. 29.0%) — 5.1 closes a gap rather than delivering an across-the-board lift. The business-workflow line tells the same story: 31.4% versus 17.1%. Mythos 5.1 separately reports 60.9% on Terminal-Bench 4.0.
Compare that with the routine tasks: CursorBench moves from 70.5% to 73.4%, a gain of just 2.9 points; Humanity's Last Exam with tools gains 1.2 points; and knowledge work sits at 1853 against Opus 5's 1824, a 29-point gap. That lines up with the pricing section above — the extra cost buys long-horizon, high-difficulty tasks, while the gap on everyday coding and routine Q&A is small enough that it may not be worth the price. Anthropic's own advice points the same way: most workloads should still start with Opus 5, with Fable 5.1 reserved for hard reasoning and long-horizon agents. The price gap backs that up — Opus 5 is $5 and $25, and it's faster too.
Sources: Anthropic model documentation, CocoLoop, release notes page; pricing, context specs, and breaking changes checked against official documentation disclosures, long-session cost estimates are rough calculations at published rates.