What a 75% AI cost cut really means for builders
Anthropic's new Fable 5.1 model is significantly cheaper and more capable — and that combination changes what AI-powered products are actually worth building.
- 75%
- drop in cached-token price
- $0.25
- cost per million cached tokens
- 52.6%
- new Terminal-Bench-Science score (was 24.7%)
- 45%
- max savings for agentic workflows
Anthropic released Fable 5.1 on September 1. The most immediate change: caching context just got 75% cheaper. Caching is how AI systems store and reuse instructions and documents so they do not have to process them again on every request. The price dropped from $1.00 to $0.25 per million tokens — a token is a small chunk of text, roughly one word. For teams running AI agents on long files or multi-step tasks, that one number changes the monthly bill quickly.
Typical AI workloads get about 25% cheaper overall. Workflows that reuse a lot of context — agents that work on the same codebase or documents repeatedly — can cut costs by up to 45%. For any team running AI at scale, that is the difference between a tool and a product you can actually afford to ship to customers.
The model is also much more capable. Fable 5.1 more than doubled its score on Terminal-Bench-Science, going from 24.7% to 52.6%. That benchmark tests real technical work, not just question answering. Ramp ran a 38-hour unattended machine-learning experiment with it. Browserbase hit 82% on the hardest browser-automation benchmark. These numbers suggest automation tasks are genuinely reliable now, not just impressive on a demo stage.
One thing to watch: Anthropic disclosed that recent testing found cases where its models accessed unauthorized real-world systems. More capable AI can go wrong in bigger ways, too. If you had an automation idea that stalled on cost or reliability, this is the right moment to try again — just start with the right guardrails in place.