# Claude Fable 5.1 cuts the cache price 75%. At max effort, a task still costs more.

URL: https://www.thedeepfeed.ai/posts/2026-09-02-claude-fable-and-mythos-5-1/
Category: Models
Published: 2026-09-02
Author: the-deep-feed
Tags: anthropic, claude, fable, mythos, pricing, benchmarks
Kind: news

> Anthropic shipped Claude Fable 5.1 to everyone and the same weights as Mythos 5.1 to vetted US organizations, cutting cache reads 75% to $0.25, but the first independent run found a max-effort task costs 20% more than on Fable 5.

## TL;DR

- **Claude Fable 5.1** and **Claude Mythos 5.1** are one model with two safeguard settings: Fable is generally available, while Mythos is limited to vetted cyber and life-sciences users, and for now only a set of **US organizations**.
- List prices stay at **$10** and **$50** per million input and output tokens; cache reads fall **75%** to **$0.25**, which Anthropic says cuts typical workload costs by about **25%** and highly agentic ones by up to **45%**.
- In Anthropic's own tables Fable 5.1 scores **52.6%** on Terminal-Bench-Science 0.1, against 24.7% for Fable 5; **Artificial Analysis** scores it **66** on its index but found a max-effort task costs **$3.76**, 20% more than Fable 5.

On September 1, **Anthropic** released [Claude Fable 5.1 and Claude Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1) and, in its own announcement, called them "the same model, but with different levels of safeguards." Fable 5.1 is generally available. Mythos 5.1 goes only to vetted cyberdefenders and life scientists, and Anthropic says "Currently, it is only available to a set of US organizations." The price story is a cache story: input and output rates stay at $10 and $50 per million tokens, while cache reads drop 75% to $0.25. Anthropic presents that as a cost cut of about 25% on typical workloads. The first independent measurement, published the same day, found the opposite at the top setting: a maximum-effort task costs 20% more than it did on Fable 5, because the new model writes far more tokens. For buyers, the effort setting now moves the bill more than the price sheet does.

> We're introducing Claude Fable 5.1 and Claude Mythos 5.1.
>
> They're the world's most advanced models for coding and knowledge work.
>
> — [@claudeai](https://x.com/claudeai/status/2094848572143407483), Sep 1, 2026

# One model, two doors

The split is the release's structural choice. According to the [developer documentation](https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1), the two models share specifications and pricing: a 1 million token context window, 128,000 output tokens, adaptive thinking that cannot be switched off, and a June 2026 knowledge cutoff. Fable 5.1 runs on the Claude API as `claude-fable-5-1` and on Amazon Bedrock, Google Cloud and Microsoft Foundry.

Mythos 5.1 is gated, and the two Anthropic documents describe the gate differently. The launch post routes access through two programs. The Life Sciences Verification Program, built "in partnership with the US government," has enrolled its first participants. The Cyber Verification Program, which today covers Opus- and Sonnet-class models with reduced cyber safeguards, "will also include access to Claude Mythos-class models" in "the near future." The developer docs say Mythos 5.1 "is offered only to approved customers in Project Glasswing," through an Anthropic, AWS or Google Cloud account team. Either way the most permissive version of the model is not something a developer can sign up for. Anthropic's Claude Security scanning product now runs on it.

That is a narrower door than the last generation got. In June, Mythos 5 and Fable 5 were pulled for everyone after [a Commerce Department letter](/posts/2026-06-27-government-joined-the-model-release/) ordered Anthropic to deny foreign nationals access. The 5.1 launch starts Mythos with US organizations only, and says Anthropic is "coordinating with the US government" to widen access to domestic and international partners.

# What changed from Fable 5

The [Fable 5.1 overview](https://platform.claude.com/docs/en/models/fable-5-1/overview) lists the full price card. Base input is $10 and output $50 per million tokens. Five-minute cache writes cost $12.50, one-hour writes $20, and cache reads $0.25, down from $1. Batch processing halves the input and output rates. The docs put cache reads at 0.025 times base input on these models, against 0.1 on other Claude models. Anthropic estimates the cut lowers typical costs by about 25% and highly agentic work by up to about 45%, based on four weeks of August usage.

The safeguards changed as much as the price. Anthropic says Claude Code users should see about 60% fewer cyber-safeguard interventions per session than on Fable 5, because Fable 5.1 may now be used to find software vulnerabilities, "though not to develop exploits for them." Penetration testing, exploit generation and binary vulnerability scanning are still redirected to Opus models. Biology safeguards fire 85% less often on benign medical and elementary biology questions than at Fable 5's launch. A new system, Enterprise Frontier Safeguards, will keep customer data in customer-controlled cloud infrastructure starting "later this fall." Until then, eligible customers can use Fable 5.1 with zero data retention.

Three changes will break existing code. Forced tool use returns an error, earlier models cannot read Fable 5.1's thinking blocks, and editing earlier turns invalidates those blocks. The last is an anti-distillation measure: Anthropic says new API accounts can no longer edit Claude's prior context while preserving its thinking transcript, a "publicly documented distillation technique." The check is enforced for accounts created on or after August 31. Fable 5.1's output also carries Anthropic's statistical text watermark, which the company says it added to comply with the EU AI Act's code of practice on AI-generated content. Effort now defaults to High in Claude Code and to Medium in Claude Cowork and on Claude.ai.

Against its own predecessor, Mythos 5, Anthropic's safety summary reports gains in both capability and alignment. Its biology capabilities exceed Mythos 5's but, by Anthropic's evaluation, still fall short of the next tier in its Responsible Scaling Policy, so it ships with the same restrictions. Anthropic says it "demonstrates the strongest cyber capabilities of any model we've released," while staying within the lower risk category of its Frontier Compliance Framework. Its automated audit found Mythos 5.1 better aligned than Mythos 5 on most metrics and less prone to reward hacking, though the model "can still sometimes bypass approvals and auto-mode classifiers."

# Anthropic's numbers, and the first outside run

Every figure below is vendor-reported, run by Anthropic with its production safeguards switched on. Where those safeguards intervened, cyber tasks were completed by Claude Opus 4.8 and biology tasks by Opus 5, which Anthropic says "likely reduces" Fable 5.1's scores.

| Benchmark (vendor-reported) | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| GDPval-AA v2 (Elo) | 1853 | 1723 | 1824 | 1711 |
| Humanity's Last Exam, no tools | 60.9% | 57.8% | 56.6% | n/a |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |

The science benchmark carries most of the launch. Simon Willison [noted](https://simonwillison.net/2026/Sep/1/claude-fable-5-1/) that it was "first announced on August 27th" and that the other gains were slight next to it; Anthropic's own footnote puts its standard error at 3.5 to 4.5 points per model. On Terminal-Bench 4.0 Anthropic also prints 60.9% for Mythos 5.1. One reader on the [Hacker News launch thread](https://news.ycombinator.com/item?id=49525378) asked why identical models score differently: ["Then why does it have separate datapoints for Terminal Bench, and score higher? Something doesn't add up here??"](https://news.ycombinator.com/item?id=49526050) Anthropic's chart caption answers it. The gap "reflects the tasks on which our earlier, less precise cyber safeguards intervened," so it measures the safeguards, not the weights.

Anthropic's scientific claims are also self-reported. It says Mythos 5.1 designed protein binders with a hit rate near 50% across 12 targets, against a typical 10 to 15%, with designs sent to two outside organizations for lab validation.

Artificial Analysis, which says it supported Anthropic with pre-release evaluation, gave the first outside read on launch day:

> Claude Fable 5.1 tops the Artificial Analysis Intelligence Index but costs 20% more per task than Fable 5 despite a 75% cache read price cut
>
> We supported @AnthropicAI with pre-release evaluation of Claude Fable 5.1. At max effort it scores 66 on the Artificial Analysis Intelligence Index, the highest score we have measured, ahead of Claude Opus 5 (max, 63), Claude Fable 5 (max, 62), GPT-5.6 Sol (max, 61) and Grok 4.6 (high, 61).
>
> — [@ArtificialAnlys](https://x.com/ArtificialAnlys/status/2094881171066978525), Sep 1, 2026

Its [write-up](https://artificialanalysis.ai/articles/claude-fable-5-1) explains the cost result. At maximum effort Fable 5.1 used about 1.7 times Fable 5's output tokens, so a task cost $3.76 against $3.14. The cache cut saved about $1.40 a task; without it the figure would have been about $5.16. At the next setting down, xhigh, the model scored 65 for $2.72, still above Opus 5's $2.34 at max. Willison's informal test showed the same spread: one prompt cost about 10 cents at low effort and $3.30 at max.

That pattern frames the release. Anthropic cut the price of re-reading context, which is what long agent sessions mostly do, and on that workload its 25% to 45% estimate may hold. It did not cut the price of thinking, and Fable 5.1 thinks more when asked to. Anthropic's own docs still tell most customers to "start with Claude Opus 5," the model [launched in July at half Fable 5's cost](/posts/2026-07-24-half-the-price-of-frontier/), and to move up only when Opus 5 at higher effort falls short. For teams that do move up, the deciding setting is the effort level they choose, not the list price.

## Sources

- [Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 (Sep 1, 2026)](https://www.anthropic.com/claude-fable-and-mythos-5-1)
- [Claude Platform Docs — What's new in Claude Fable 5.1 (Sep 1, 2026)](https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1)
- [Artificial Analysis — Claude Fable 5.1 tops the Artificial Analysis Intelligence Index but costs 20% more per task than Fable 5 (Sep 1, 2026)](https://artificialanalysis.ai/articles/claude-fable-5-1)
- [Simon Willison — Claude Fable 5.1 made me a really nice animated pelican (Sep 1, 2026)](https://simonwillison.net/2026/Sep/1/claude-fable-5-1/)
- [Hacker News — Claude Fable 5.1 and Claude Mythos 5.1 (1,419 points, 1,395 comments)](https://news.ycombinator.com/item?id=49525378)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-09-02-claude-fable-and-mythos-5-1/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt