Claude Fable 5.1 Kept the Same Sticker Price but Bills 25% Less: The Saving Is All in Cache Reads, Down 75% (Checked Sept 2026)

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 2026-09-01.

Here is the odd part: US$10 per million input tokens and US$50 per million output tokens — exactly the same sticker price as Fable 5. Nothing came down. Yet the announcement says typical workloads cost about 25% less, and heavily agentic ones up to about 45% less.

If the sticker price did not move, where does the money come from? From a line most people have never looked at: cache reads.

1. The line that actually got cheaper

Anthropic’s announcement states it plainly: Fable 5.1 cache reads cost US$0.25 per million tokens, 75% less than Fable 5.

Checking that against the official docs, which express cache pricing as a multiplier:

• Most models: cache read = 0.1x the base input price
• Claude Fable 5.1 and Mythos 5.1: 0.025x

US$10 × 0.025 = US$0.25. The two sources agree.

So the same repeated chunk of context that cost US$1 to re-read on Fable 5 now costs US$0.25.

2. Why this hits agentic work hardest

What is a cache read? In plain terms: the big unchanging block at the front of every request — your tool definitions, your system instructions, the document you pasted — gets charged once, and every later request that reuses it pays the discounted rate.

That explains why the announcement quotes two different numbers:

• ~25% on typical workloads: one-off questions, not much repetition.
• up to ~45% on heavily agentic work: an agent running fifteen or twenty steps re-reads the same tool list and the same document every single step. Repetition dominates the bill, so a cache discount is amplified.

Put simply: the longer and more multi-step the job, the more this price change is worth to you.

Worth stating clearly: the 25% / 45% figures are Anthropic’s description of typical cases, not a guarantee. Your own saving depends on how much of your prompt repeats — and on caching being configured correctly in the first place. We cover that in How to set up Claude prompt caching so it actually saves money.

3. Fable 5.1 vs Mythos 5.1

Two names shipped at once, and Anthropic is direct about it: they are the same model with different levels of safeguards.

• Claude Fable 5.1: generally available, model id claude-fable-5-1.
• Claude Mythos 5.1: available only through trusted access programs, currently limited to US organisations, via the Cybersecurity Verification Program (CVP) and Life Sciences Verification Program (LSVP), with safeguards designed to support work in those fields.

Anthropic also notes that the newest cybersecurity safeguards produce 60% fewer false positives than before — legitimate security work gets blocked less often.

4. A new limit that is easy to miss

One item in the announcement has nothing to do with price but will trip up developers:

To prevent distillation attacks, new API accounts cannot manually edit Claude’s prior context in a multi-turn conversation.

If your workflow rewrites the model’s previous reply before sending it back, that will not work on a new account. It also interacts with caching: cache invalidation is hierarchical (tools → system → messages), so editing earlier content invalidates everything downstream and wastes the money you already spent writing that cache.

5. Where you can use it

Anthropic lists Fable 5.1 as generally available on:

• Claude API
• Amazon Web Services
• Google Cloud
• Microsoft Azure
• Claude Code, Claude Enterprise, Claude Platform

Mythos 5.1 remains limited to the trusted access programs described above.

Separately, GitHub’s 2026-09-04 changelog announced Claude Fable 5.1 for Copilot Pro+, Max, Business and Enterprise users — we covered the data retention side of that in Claude Fable 5.1 in Copilot and what happens to your code.

6. Three-line summary

1. Fable 5.1’s input and output prices did not change (US$10 / US$50). Cache reads did: US$1 → US$0.25.
2. Anthropic quotes ~25% typical and up to 45% on agentic work — more repetition means more saving, but only if caching is set up correctly.
3. New API accounts cannot hand-edit the model’s earlier replies, a new anti-distillation limit shipped alongside.

Official links

• Announcement: https://www.anthropic.com/claude-fable-and-mythos-5-1
• Prompt caching docs: https://platform.claude.com/docs/en/build-with-claude/prompt-caching

More on this site

• Claude tool page
• How to set up Claude prompt caching so it actually saves money
• Claude Fable 5.1 in Copilot and what happens to your code
• Claude Sonnet 5 API price lock

All figures were read directly from the official announcement and docs on 2026-09-05. Prices and policies change; check the official pages before you commit.