Claude Fable 5.1: 75% cheaper cache reads, but three API behaviors change
Anthropic's Claude Fable 5.1 release has an interesting tradeoff for developers building long-running agents.
Cache reads are now $0.25 per million tokens, down 75% from Fable 5.
That sounds like an easy upgrade.
It isn't.
Fable 5.1 changes several behaviors that can break existing agent implementations:
tool_choice: any and forced named tools now return errors
Thinking blocks have directional compatibility between model generations
Editing earlier context around thinking blocks can trigger validation errors
The model can make more individual tool calls
Whole-file rewrites can increase output-token consumption
A 30-day retention requirement applies unless zero-data-retention access is authorized
This makes the economics more interesting.
A long-running coding agent may have a huge cached context, but if the new model produces more tool calls or larger outputs, some of the cache savings can disappear.
The right metric is therefore not simply:
cost per million tokens
but:
cost per accepted task
That means measuring cache hits, retries, tool calls, output volume, fallbacks, latency, and human review.
I've broken down the changes and the migration checklist in the full article.
🔗 Read the full article
ai #claude #anthropic #llm #agents #coding #developers #machinelearning
Medium
Top comments (0)