Introduction
Anthropic released Claude Fable 5.1 and the invitation-only Claude Mythos 5.1 on September 1, 2026. If you compare the two price sheets side by side, nothing happened: input is $10 per million tokens and output is $50 per million, exactly what Claude Fable 5 charged in June.
The change is one line further down. A prompt cache read on Fable 5.1 costs $0.25 per million tokens, down from $1 — a 75% cut, and the only price that moved. Anthropic says that lowers a typical bill by roughly 25% versus Fable 5, and by up to 45% for highly agentic work, based on four weeks of its own August 2026 usage across Claude Enterprise, Claude Code and the API. That is a release aimed at one specific workload, and a reader scanning headline rates would miss it entirely.
The Cache Read Is the Release
Every other Claude model prices a cache hit at 10% of base input. Fable 5.1 and Mythos 5.1 price it at 2.5%. Anthropic's own model documentation states the exception in a footnote: "prompt caching reads cost 10% of the base input price (2.5% on Claude Fable 5.1 and Claude Mythos 5.1)."
Why that is worth more than a sticker-price cut depends on the shape of the workload. An agent does not send a fresh prompt each turn; it replays a growing conversation, a tool schema and a pile of retrieved files, then appends one new instruction. Cached tokens dominate, and they dominate more the longer the session runs.
| Item | Claude Fable 5.1 | Claude Fable 5 |
|---|---|---|
| Input | $10 / MTok | $10 / MTok |
| Output | $50 / MTok | $50 / MTok |
| 5-minute cache write | $12.50 / MTok | $12.50 / MTok |
| 1-hour cache write | $20 / MTok | $20 / MTok |
| Cache read | $0.25 / MTok | $1 / MTok |
Put numbers on it. A coding agent holding a 200,000-token cached prefix across 50 turns reads 10 million cached tokens. On Fable 5 that is $10 in cache reads alone; on Fable 5.1 it is $2.50. Nothing else in the request changed, and no prompt was rewritten to get it.
Two caveats belong next to the claim. The 25% and 45% figures come from Anthropic's own traffic, not an independent study, and Anthropic chose which workloads count as "typical" and which as "highly agentic." And the discount only reaches requests that actually hit a warm cache: a single-shot call, or a session that rebuilds its prefix each turn, pays exactly what Fable 5 paid. The minimum cacheable prompt is still 512 tokens, so the break-even now favors caching almost everything.
What Else Changed
Fable 5.1 is a point release, and most of its specifications carry over: the API model ID is claude-fable-5-1, the context window is 1M tokens at standard per-token pricing across the whole range, max output is 128K tokens per synchronous request, and adaptive thinking is always on and cannot be disabled.
Two things did move:
- The knowledge cutoff is now June 2026, up from January 2026 on Fable 5 — five months of fresher training data, and the widest cutoff gain in the Claude lineup this year.
- Effort defaults now differ by surface. The five levels are
low,medium,high,xhighandmax. The Claude API defaults tohigh, and so does Claude Code — but Claude Cowork and Claude.ai default tomedium. If you benchmark the model in one product and deploy it through another, you are not comparing the same configuration.
Three changes are breaking for anyone already calling Fable 5: forced tool use returns a 400, thinking blocks cannot be read by earlier models, and editing anything before a thinking block invalidates every later one. The full list is on our Claude Fable 5.1 model page.
The Benchmark Table Is Text This Time
Anthropic published the 5.1 results as a text table. That sounds like nothing. It is a direct fix to the complaint our own model page raised in June: the Fable 5 launch post carried its comparison table only as an image, and not one score from it appeared anywhere in the post's prose. Transcribing a vendor's headline numbers out of a PNG is not something a reader, a screen reader or an answer engine should have to do. Crediting the correction is fair.
The numbers themselves are vendor-run. Anthropic chose the benchmarks, the harnesses and the effort levels, and ran them itself.
| Benchmark | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 |
|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% |
| GDPval-AA v2 | 1853 | 1723 | 1824 |
| OSWorld 2.0, partial credit | 77.9% | 72.9% | 75.4% |
| OSWorld 2.0, strict | 41.7% | 36.1% | 39.6% |
| Humanity's Last Exam, no tools | 60.9% | 57.8% | 56.6% |
| Humanity's Last Exam, with tools | 65.0% | 63.8% | 63.6% |
Claude Mythos 5.1 scores 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8% — the same model, five points apart, with the difference being production safeguards. It is the clearest published price tag on those safeguards Anthropic has offered.
One footnote deserves a second read: Fable was evaluated with safeguards enabled, and where a safeguard intervened, cybersecurity tasks were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5. Part of what these rows measure is a different model finishing the job.
Mythos 5.1 and Who Actually Gets It
Claude Mythos 5.1 (claude-mythos-5-1) shares Fable 5.1's specifications and pricing with some safeguards relaxed for professional cybersecurity and life-sciences work. Access is invitation-only, and Anthropic states plainly that "currently, it is only available to a set of US organizations."
The routing is worth reading carefully, because it is easy to overstate. Project Glasswing was not replaced — Anthropic's model docs still describe Mythos 5.1 as offered "by invitation only, as part of Project Glasswing," with access arranged through an Anthropic, AWS or Google Cloud account team, and Glasswing is itself being extended to roughly 150 new organizations in more than fifteen countries. The two verification programs are the professional routes running under that umbrella: the Life Sciences Verification Program, an invite-only beta run with the US government, and the Cyber Verification Program, which today grants reduced cyber safeguards on Opus- and Sonnet-class models. Anthropic says the CVP "will also include access to Claude Mythos-class models" in the near future — that is a plan, not a live entitlement.
Also unchanged from Fable 5: both models are Covered Models with a mandatory 30-day retention window, so neither is available under zero data retention without explicit authorization. And the June 2026 export-control episode, when Anthropic suspended access to Fable 5 and Mythos 5 outright for nearly three weeks, goes unmentioned in the launch materials. It is still the reason to keep a fallback model configured.
Where It Lands in a Four-Day Frontier Shuffle
The frontier moved three times in four days: Fable 5.1 on September 1, Gemini 3.8 Flash on September 2, and GPT-6 Astra on September 3-4.
Cross-vendor comparison is harder than usual right now, because Artificial Analysis shipped Intelligence Index v4.2 on September 4 and its scores do not carry over from v4.1. On v4.2, Artificial Analysis places Fable 5.1 at 57, GPT-6 Astra at 55 and Gemini 3.8 Flash at 59. Any v4.1 number you find for these models — including Fable 5's 60 — is measuring a different exam.
Note also that GPT-6 Astra matches Fable 5.1's $10/$50 exactly, so the two flagships now share a sticker price and differ on the terms around it. Astra bills prompts over 272K input tokens at 2x input and 1.5x output for the entire request; Fable 5.1 charges standard rates across its full 1M window and discounts the replay. Which is cheaper is a question about your context length and cache hit rate, not about the rate card.
Conclusion
Fable 5.1 is a modest capability release wrapped around a targeted price change. The benchmark gains over Fable 5 are real but incremental outside Terminal-Bench-Science, where the jump from 24.7% to 52.6% is genuinely large. The cache read cut is the part that changes decisions: it makes long-running agents on Anthropic's most expensive model materially cheaper without touching a single prompt, and it makes the model no cheaper at all for one-shot calls.
Check your cache hit rate before you assume the 25% applies to you. And if you cite a benchmark score from any of this week's launches, cite the index version with it.