An agent reads the same specification again. Then the same repository context. Then both again, after a test fails. That repeated input is where Claude Fable 5.1 changes the bill.
Anthropic's September 1 release keeps Fable's base rates at $10 per million input tokens and $50 per million output tokens. Cached reads fall from $1 to $0.25 per million tokens. A request with a million already-cached input tokens therefore saves $0.75 on that part of the request. Output and cache writes still cost extra. Claude pricing.
That distinction matters. A 75% reduction in the cache-read rate is not a 75% reduction in your total bill. A short request with no reusable prefix gets none of that saving. A long agent loop may get quite a lot.
Anthropic positions the model for harder coding, research, and work across documents, spreadsheets, and slides. The API retains a million-token context, supports up to 128K output tokens, and uses always-on adaptive thinking. Its own guidance is useful: evaluate Fable 5.1 when a higher-effort Opus 5 run still falls short. Model documentation.
The release also separates two access paths. Fable 5.1 is generally available; Mythos 5.1 uses the same underlying model with different safeguards and remains restricted to trusted access programs. Anthropic says Fable can now help discover software vulnerabilities, while exploit development remains restricted. That is a change from treating all cyber work as a reason to fall back to an older model. Release announcement.
For a useful trial, keep the task and acceptance checks fixed. Record cache hits, output tokens, retries, and whether the final artifact works. The model tools include Fable 5.1 now; local scenario results remain marked “not tested” until there is an actual run to compare.