Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, 2026. Fable 5.1 is generally available today across Claude.ai, the API, AWS, Google Cloud, and Microsoft Azure, under the model id claude-fable-5-1. Headline input and output pricing did not move: $10 per million input tokens and $50 per million output. What moved is the number almost nobody puts on a slide. Cache reads dropped from $1.00 per million tokens to $0.25, a 75 percent cut.
That single line is why Anthropic estimates roughly 25 percent lower cost for typical workloads and up to about 45 percent for heavily agentic ones. If you run long-context agents, the cache read is the line item you were actually paying.
What actually changed in the pricing
Most Claude models price a cache read at 10 percent of the input rate, as the published rate card shows. Fable 5.1 prices it at 2.5 percent. Against a $10 input rate, that is $0.25 per million rather than $1.00.
| Token type | Fable 5 | Fable 5.1 | Change |
|---|---|---|---|
| Input | $10 | $10 | unchanged |
| Output | $50 | $50 | unchanged |
| Cache read | $1.00 | $0.25 | 75 percent lower |
The reason this matters more than a headline cut is structural. An agent that reads a large codebase, a document set, or a long conversation history re-reads the same cached context on every turn. Those reads dominate the bill on long-running tasks, which is exactly why the saving climbs to around 45 percent on agentic work and sits near 25 percent on ordinary use.
It is a different shape of price move from the one OpenAI made on GPT-5.6 Sol last month, which cut the headline output rate. Anthropic left the headline alone and cut the rate you hit repeatedly. Which of those saves you more depends entirely on whether your workload is output-heavy or context-heavy.
The benchmark gains are not marginal
Anthropic published side-by-side numbers against Fable 5, and several of them are unusually large for a point release.
| Benchmark | Fable 5 | Fable 5.1 |
|---|---|---|
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% |
| Terminal-Bench 4.0 | 42.0% | 55.8% |
| AutomationBench | 17.1% | 31.4% |
| OSWorld 2.0 (partial) | 72.9% | 77.9% |
| CursorBench 3.2.0 | 70.5% | 73.4% |
Terminal-Bench-Science more than doubles and AutomationBench nearly does. Those are both long-horizon, tool-using benchmarks, which lines up with Anthropic's framing that the model is better at avoiding shortcuts and at working toward the underlying cause of a problem rather than patching a symptom. CursorBench moves least, which is a useful reality check: single-file editing was already strong and this release is not aimed there.
Why the cache line is the one that moves bills
Prompt caching works by letting you mark a stable chunk of context, a system prompt, a schema, a repository snapshot, so the model does not re-process it from scratch on every call. Anthropic documents the mechanics in its prompt caching guide, and the current rate card sits in the pricing docs.
The economics follow from how often each token type is touched. A single agent task might write the cache once, emit a few thousand output tokens, and read the same cached block on every one of twenty turns. Under the old 10 percent multiplier those repeated reads quietly became the largest line on the invoice. At 2.5 percent they stop being the thing you optimise around.
That is why the saving is a range rather than a number. A one-shot call with a short prompt sees almost nothing. A long agent loop over a large cached context sees close to the top of the range, which is where the 45 percent figure comes from.
What this changes for builders
Re-price your agent loops before you re-architect them. Teams have spent the past year trimming context to control cost. If your prompt caching was already working, a 75 percent cut on reads may make aggressive context trimming unnecessary, and trimming usually costs accuracy.
Long-context work becomes the cheap path, not the expensive one. Feeding a whole repository or asset library and letting the cache carry it across turns is now priced closer to how people actually want to work.
Output-heavy jobs see nothing. Output stayed at $50 per million. If your pipeline generates long text and reads little context, this release does not lower your bill and you should evaluate it on capability alone.
The timing is also convenient for anyone forced to move: OpenAI is winding down Cursor's access to its models with a November 12 cutoff, so a chunk of the editor market is shopping for a replacement this quarter.
Checking whether the saving is real for you
1. Pull your cache read volume, not your total tokens. The saving is proportional to cached reads only. If your dashboard does not separate them, that is the first thing to fix.
2. Recompute at $0.25 per million. Multiply last month's cache reads by the new rate and compare. This tells you whether you are in the 25 percent case or nearer the 45 percent one.
3. Re-test a job you previously trimmed. Take a workload where you cut context to save money, restore the full context, and score accuracy against the trimmed version. If accuracy improves, the trim was costing you quality that is now cheap to recover.
4. Leave output-heavy jobs alone. They are unaffected. Judge them on the benchmark gains instead, particularly if they involve tool use.
What else shipped alongside
Anthropic reports 60 percent fewer false positives in the cybersecurity safeguards and fewer false refusals on benign biology and medical queries, along with strengthened anti-distillation measures. Outputs also carry invisible watermarks for EU AI Act compliance, with a detection API in private preview.
On the science side, Anthropic cites protein binders designed with 10 times higher affinity than the competition, a high-resolution Venus elevation map built from NASA Magellan data and published openly on Zenodo, and GPU kernels for genomic models optimised up to 2.5 times faster. The system card carries the evaluation detail.
The safeguard work has its own write-ups, covering the biology safeguards and expanded support for scientists, and 9to5Mac's launch-day coverage is a shorter summary of the consumer-facing changes.
A note on specifications: Anthropic's announcement does not publish a context window or maximum output figure for 5.1, and several secondary write-ups quote numbers that do not appear in the primary source. Treat any specific context or output ceiling you see elsewhere as unconfirmed until Anthropic states it. We will update this piece if Anthropic publishes them.
Frequently asked questions
How much cheaper is Claude Fable 5.1?
Headline input and output rates are unchanged at $10 and $50 per million tokens. Cache reads fell from $1.00 to $0.25 per million, a 75 percent cut, which Anthropic estimates at roughly 25 percent lower total cost for typical workloads and up to about 45 percent for agentic ones.
Why does a cache read discount matter so much?
Agents that work over a large codebase or document set re-read the same cached context on every turn, so cache reads dominate long-running bills. Cutting that rate reduces the line item you hit most often rather than the one you quote most often.
Is Fable 5.1 available to everyone?
Yes. It is generally available on Claude.ai, the API, AWS, Google Cloud, and Microsoft Azure, under the model id claude-fable-5-1. Mythos 5.1 is the restricted sibling and is not generally available.
What improved most in the benchmarks?
Long-horizon tool use. Terminal-Bench-Science 0.1 went from 24.7 to 52.6 percent and AutomationBench from 17.1 to 31.4 percent. CursorBench moved least, from 70.5 to 73.4 percent, so single-file editing is not where this release concentrates.
What is the context window on Fable 5.1?
Anthropic has not published one for 5.1. Figures circulating in secondary coverage do not appear in the primary announcement, so treat them as unconfirmed.
Does this help if my workload is output-heavy?
No. Output pricing is unchanged at $50 per million. Evaluate 5.1 on capability rather than cost if you generate a lot of text and read little cached context.
