AI Agents Just Got 75% Cheaper to Keep Running
Anthropic cut the cost of cache reads — the most repeated operation in long-running AI agents — by 75%, which means the monthly bill for teams automating real work just dropped meaningfully.
- 75%
- drop in cache-read token price
- $0.25
- new price per million cached tokens
- 45%
- max bill reduction for cache-heavy agents
- 52.6%
- Fable 5.1 score on Terminal-Bench-Science
The most common cost in running an AI agent is not the first answer — it's every answer after that. When an agent works through a long task, it re-reads the same background context over and over. Those re-reads are called cache reads — the little chunks of stored context an AI charges you for — and Anthropic cut their price by 75%, from $1.00 to $0.25 per million tokens.
To understand why this matters, think of a customer-support agent that reads a 50,000-word product manual before every reply. Before this change, that re-reading added up fast. With the new price, the same agent costs four times less to run in the background. That directly changes what is affordable.
Anthropic estimates that for most workloads the total bill drops around 25%, and for agents that rely heavily on cached context — research tools, code assistants, document processors — savings can reach 45%. The new model, Fable 5.1, also doubled its score on agentic benchmarks compared to the previous version: from 24.7% to 52.6% on the Terminal-Bench-Science test.
If your team is running or planning to run AI agents at any real scale, this is the moment to re-check your cost model. Work that seemed too expensive six months ago may now be viable. The price of AI computation keeps falling — and this cut targets exactly the kind of repeated-context work that real agents do most.