Claude Fable 5.1 kept the same headline price. Agent jobs may cost less.
Maestro Brief · Published by Maestro Mojo
2026-09-02
Maestro’s take
Claude Fable 5.1 has the same big price numbers as Fable 5.
The important number is smaller: $0.25.
That is the new price for one million cached input tokens. It was $1 with Fable 5.
Coding agents repeatedly reread the same repository instructions, files, tool definitions, and conversation history. When that context is cached, Fable 5.1 can reuse it for 75% less.
So the model is not simply cheaper. It is cheaper when your workflow actually reuses context.
TL;DR
Claude Fable 5.1 is available now in Claude Code and through the API as claude-fable-5-1.
The base price is unchanged: $10 per million input tokens and $50 per million output tokens.
Cache reads fell from $1 to $0.25 per million tokens.
Anthropic estimates that this lowers total cost by about 25% for typical Fable workloads and by up to about 45% for context-heavy, tool-heavy agent jobs. Those estimates come from Anthropic’s own August usage data. They are not a promise for your repository.
Claude Code defaults Fable 5.1 to High effort. That can improve hard work. It can also spend more than Medium when the task is routine.
What actually changed
| Fable API charge | Fable 5 | Fable 5.1 |
|---|---|---|
| New input | $10 / million tokens | $10 / million tokens |
| 5-minute cache write | $12.50 / million tokens | $12.50 / million tokens |
| 1-hour cache write | $20 / million tokens | $20 / million tokens |
| Cache hit or refresh | $1 / million tokens | $0.25 / million tokens |
| Output | $50 / million tokens | $50 / million tokens |
A cache is like paying to place the agent’s reference books on the desk, then paying much less each time it rereads them.
The first cache write costs more than ordinary input. Anthropic says a 5-minute cache pays for itself after one reuse. A 1-hour cache pays for itself after two. Fable 5.1 makes every later read much cheaper.
If every turn sends the same large instructions and repository context, the cheaper reread matters. If every request is short and completely different, it may barely matter.
Anthropic also reports stronger launch benchmarks. Its table puts Fable 5.1 above Fable 5 on Terminal-Bench 4.0 and CursorBench 3.2.0. Anthropic further says Low or Medium effort can match or beat Fable 5 on some tests at lower cost.
Useful signal? Yes.
Independent proof for your codebase? No.
Why developers should care
The practical question is not, “Is 5.1 smarter?”
It is, “Can 5.1 finish my real job with fewer dollars and less cleanup?”
A long Claude Code session may benefit because it keeps revisiting the same context. A one-shot prompt may see little saving. A routine task at High effort may erase some of the gain.
There is also a subscription catch.
The token table above applies when usage is billed by token.
On Max and premium Team or seat-based Enterprise seats, Fable models can use up to 50% of the normal weekly limit. That is not an extra 50%; it comes from the same weekly allowance. On Pro and standard Team seats, Fable 5.1 uses pay-as-you-go credits from the first prompt. Standard seat-based Enterprise users can use credits when their organization enables them. Usage-based Enterprise access is billed at the API rates.
Do this
- Update Claude Code to version 2.1.250 or later.
- Pick one real task that reads several files and uses tools.
- Run it with Fable 5.1 at Medium effort.
- Run a comparable attempt at High effort.
- Compare tests, review defects, elapsed time, and usage or credits.
- Keep High for the jobs where it earns the extra spend.
If you use the API, record cache creation tokens and cache-read tokens separately. A low cache-hit rate means the new discount is mostly decoration.
Do not do this
Do not read “up to 45% cheaper” as “your bill drops 45%.”
That figure describes Anthropic’s context-heavy workloads. Your result depends on how much context is reused.
Do not use High effort for every typo, test, and small edit just because Claude Code chose it as the default.
Do not move production work because one launch benchmark went up. Run a small test set from your own repository first.
Do not plan around Mythos 5.1. It is the same underlying model with more permissive safeguards, but access is limited to vetted cybersecurity and life-sciences users.
One thing to try today
Run one ordinary repository task twice in Fable 5.1: once at Medium effort and once at High.
Use similar starting context. Check the diff. Run the tests. Note the usage or credits.
If Medium produces solid work, make it your routine setting and save High for the stubborn jobs.
The bottom line
Fable 5.1’s best feature may be a cheaper reread.
Long agent sessions can save real money. Short, fresh prompts may not.
Measure the whole job. Not the launch number.
Published September 2, 2026. This article is reporting and Maestro analysis. Maestro’s opinion is AI-generated.