Anthropic prices Fable 5.1 at $10/1M input tokens and $50/1M output tokens, the same as Fable 5's pricing, but cuts cache read pricing by 75% to $0.25/1M tokens
Frederic Lardinois /The New Stack:
Context & Ripple Effects
Fable 5.1 follows Fable 5's June price reset, which set $10-per-million input and $50-per-million output rates for Fable and Mythos. Anthropic had already said the new release would lower estimated workload costs by 25% for typical use and up to 45% for highly agentic work; the cache-read rate explains a material part of that claim.
The pricing change also addresses a usage problem exposed by Ramp's spending data, which showed Fable 5 plateauing as companies moved toward cheaper models. Keeping headline token rates flat while discounting reused context targets the workloads most likely to accumulate large input volumes.
First-order effects
- Anthropic customers that repeatedly reuse prompts, codebases, or agent state on Fable 5.1 pay $0.25 rather than $1 per million cached input tokens, while standard input and output prices remain unchanged.
- Fable 5.1's effective cost becomes more dependent on workload architecture: teams that can maximize cache reuse receive the largest savings rather than every request receiving a lower list rate.
Second-order effects
- Enterprise buyers comparing Fable against lower-priced alternatives gain a narrower cost gap for long-running, context-heavy deployments, the segment where Fable 5 had faced spending pressure.
- Anthropic's model-selection calculus changes inside its own lineup: commentators argued that the cache discount makes Fable cheaper than Opus for agentic workloads, making cache behavior a more important routing criterion.
Third-order effects
- Agent-model competition is increasingly moving from published input and output rates to effective inference cost, where context reuse and agent-loop design determine the bill.
- If providers continue to price cached context aggressively, application developers gain an incentive to standardize persistent context and caching layers as core infrastructure for agentic systems.
The trend: Frontier-model vendors are using workload-specific inference pricing to make agentic deployments economical without cutting their headline token rates.