Claude Fable 5.1 and Claude Mythos 5.1

TL;DR Highlight

Fable 5.1 cuts typical cost about 25% via cheaper cache reads and improves coding benchmarks.

Who Should Read

A backend developer running coding agents on the Claude API who is weighing a model switch and cost optimization.

Core Mechanics

Fable 5.1 and Mythos 5.1 are the same model
- Only the safeguard level differs.
- Fable 5.1 is generally available while Mythos 5.1 is limited to trusted access programs.
Cheaper cache reads cut the bill
- Token-billed typical workloads cost an estimated 25% less than Fable 5.
- Highly agentic work often saves more, up to about 45%.
Low or Medium effort matches Fable 5
- At Low or Medium effort it reaches results similar to or better than Fable 5 at much lower cost.
- Claude Code defaults to High, while Claude Cowork and Claude.ai default to Medium.
Benchmarks rose over Fable 5
- Terminal-Bench 4.0 is 55.8% versus 42.0% for Fable 5.
- CursorBench 3.2.0 is 73.4% versus 70.5% for Fable 5.
- Terminal-Bench-Science 0.1 is 52.6% versus 24.7% for Fable 5.

Community Reactions

Whole-file rewrites flagged
- A commenter noted the changelog says the model is more likely to rewrite whole files for small edits, costing more output tokens and time.
- The suggested fix is adding the official docs' "edit surgically" instruction to the system prompt.
Usage limits burned fast
- Commenters reported hitting the 5-hour limit before finishing a first task on both Team Premium and Max 20x accounts.
- One said a multi-hour run felt plodding and took another 1h15m just to wrap up.
Pushback on the size of the gain
- One commenter argued that without Terminal-Bench-Science the gain over Opus 5 is only 1.5 to 3.5 points.
- Another said Artificial Analysis shows Fable 5.1 costing more than 5.0.
Token budgets matter more than models
- One commenter said predictable spend within monthly quotas matters more than a better model.
- Another said cheap models like DeepSeek V4 Flash or GLM 5.3 Flash are good enough.

How to Apply

Add a surgical-edit instruction
- Append the instruction in the prompt below to your coding agent's system prompt today.
- Run the same edit task before and after and compare output tokens.
Compare cost at Low and Medium effort
- Run a few existing Fable 5 or Opus 5 tasks on Fable 5.1 at Low and Medium.
- Log cost per task and quality in a table to choose your default effort.
Migrate cache-heavy agents first
- Move workloads that reread the same context, such as code review or long-running agents.
- Check the change in cache read cost on your bill before and after.
Re-measure false positives on security tooling
- Rerun vulnerability analysis prompts that safeguards blocked before.
- Record the block rate.

Try It

copy & run

# Paste at the end of your Claude Code or API system prompt (instruction quoted from the official docs via a comment)

The number of tokens used to edit files is best minimized, all else being equal. Therefore, when it will not affect the end result, try to surgically edit a file rather than rewrite the entire thing.

# Before and after adding this, give {{file}} the same small edit {{request}} and compare output token counts.

Terminology

effortA setting for how much reasoning the model spends on a problem. Like an AC dial, higher gives better results but costs more time and money.

cache readThe model rereading input it has already processed and stored. It is like keeping a book open on your desk, so it is billed cheaper.

zero data retentionA policy where the provider does not store data sent via the API. Like a shop that throws away the receipt once the sale is done.

false positiveA benign request wrongly flagged as dangerous. Like a security scanner beeping at your keys.