Fable 5.1 Ships Today: Cache Reads Drop From $1.00 to $0.25 per Million Tokens

TL;DR: Anthropic launched Fable 5.1 and Mythos 5.1 today (September 1). Cache reads drop from $1.00 to $0.25 per million tokens — Anthropic reports this cuts effective costs 25% on typical workloads and up to 45% on heavily agentic ones.

Cache reads are the compounding cost in agentic systems: every time an agent revisits the same system prompt, tool definitions, or accumulated context, it triggers a read. Fable 5's $1.00/M rate made this expensive for long-running Claude Code sessions and multi-step pipelines. At $0.25/M, the same re-use is 75% cheaper. Cache write prices are unchanged ($12.50/M for 5-min, $20/M for 1-hr); the savings are entirely on reads.

Input and output prices are unchanged at $10/$50 per million tokens. The model outperforms Fable 5 on Anthropic's internal evals, with the biggest gains in computer use, long-horizon agentic work, and terminal-based engineering tasks. Knowledge cutoff is June 2026. Fable 5.1's cybersecurity capabilities are strong enough that some queries in that domain are automatically rerouted to Opus models — test any security tooling built on the API before assuming behavior is identical to Fable 5.

Mythos 5.1 ships alongside with restricted access to vetted US organizations (cyberdefenders and life scientists only, via a separate application process). Not broadly available.

What to do today

If you've enabled prompt caching in your pipeline (via the cache_control parameter), update your model string to the Fable 5.1 identifier (see the Anthropic announcement for the exact string). No other code changes needed — the 75% read savings apply automatically to any already-cached context.

Limits:

• The 25–45% effective cost reduction is Anthropic's own estimate; independent controlled comparisons don't exist yet. Treat it as directionally correct.
• At $10/$50/M input/output, Fable 5.1 remains the most expensive tier in the Anthropic lineup. Cache savings help pipelines with heavy context re-use; fresh-context requests see no change.
• Knowledge cutoff June 2026 means it has no awareness of events from the past three months.

Why it matters: Agentic pipelines that re-use context pay 75% less per cache read starting today — update the model string and the savings apply automatically.

Sources: Introducing Fable 5.1 and Mythos 5.1 — Anthropic · System Card, Sep 1 2026 — Anthropic · Cache reads 75% cheaper — VentureBeat · Fable 5.1: cheaper, smarter, refuses less — The New Stack