Claude Fable 5.1 pricing turns repeated context into the release, with cached input now costing 75% less. Anthropic estimates 25% lower costs for typical workloads and up to 45% for highly agentic work. The discount rewards agents that keep useful working memory instead of rebuilding it on every turn.
Why is cache pricing the headline?
Anthropic left standard input and output prices unchanged at $10 and $50 per million tokens. Cache reads fell to $0.25 per million. That makes the discount specific: it applies when the model reads context it has already processed.
The company published two estimates in its Fable 5.1 announcement. Because cache reads alone became 75% cheaper, we can work backwards from those totals.
| Workload | Estimated total saving | Implied old bill from cache reads |
|---|---|---|
| Typical | 25% | 33% |
| Highly agentic | up to 45% | up to 60% |
The arithmetic is total saving ÷ 75%. It says the expensive part of a long agent run is often not the new prompt. It is rereading the project map, tool results and earlier decisions that keep the run coherent.
On an old $100 typical-workload bill, the estimate implies about $33 went to cache reads. The new rate cuts $25 from that slice while leaving the other $67 alone.
What changes for coding agents?
Cheaper memory changes the optimization target. The first question is no longer how little context an agent can carry. It is which context earns its place on every turn.
- Keep decisions. Record constraints and rejected approaches so the agent does not reopen settled work.
- Keep receipts. Preserve test output and failure evidence so verification survives a long run.
- Drop drift. Remove stale plans and contradictory notes before they become cheap misinformation.
This is the same reason our agent-facing theme files keep tokens, rules and workflows in readable artifacts. Persistence helps only when the material being persisted is maintained.
Check one agent run
Split one completed task into uncached input, cache reads and output. If cache reads were less than one-third of the old bill, do not expect the headline 25% saving.
Where can the savings disappear?
The estimates are not a new flat price. They came from four weeks of Anthropic usage in August 2026 at default effort. Your result still depends on cache hits, output volume and the effort setting.
The tradeoff is easy to miss: cheaper context can make oversized context look harmless. A long-lived agent with an accurate notebook gets cheaper. One carrying stale instructions gets cheaper at repeating the wrong thing.
Fable 5.1 also posts stronger coding and research scores. Anthropic reports 55.8% on Terminal-Bench 4.0, up from Fable 5 at 42.0%. Those gains matter, but a benchmark does not tell you what a week of delegated work will cost.. the cache line does.