Claude Fable 5.1 and Mythos 5.1: a 75% cache discount aimed squarely at agentic coding costs

Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1 — same underlying capability, different access tiers, and a pricing change aimed squarely at one thing: making agentic coding workflows cheaper to run at scale.

What actually changed

Sticker prices for Fable 5.1 stay put at $10 per million input tokens and $50 per million output tokens. The real move is on cache reads, which drop 75% to $0.25 per million tokens. For typical usage that works out to roughly a 25% cost cut; for agentic workloads — the repeated, context-heavy calls that coding agents and long-running tool-use loops generate — Anthropic and multiple outlets reporting on the release put the savings at up to 45%.

That distinction matters. Agentic coding tools re-send large chunks of context (file trees, prior tool outputs, system prompts) on nearly every turn. Cache-read pricing is the line item that scales with that pattern, not the flat input/output rate — so cutting it 75% is a more targeted lever than a blanket price cut would have been.

Two models, one weight class

Fable 5.1 and Mythos 5.1 are, per reporting, the same model with different safeguard configurations rather than two separate weight classes. Fable 5.1 is generally available; Mythos 5.1 is restricted to organizations and individuals inside Anthropic’s trusted-access programs, primarily U.S.-based. If you’re building on the public API, Fable 5.1 is the one you’ll actually reach for.

The watermark nobody asked for (but regulators did)

This is the first Claude release to ship invisible watermarking in text and file outputs, following Anthropic’s earlier commitment to watermark models released after August 2, 2026 — timed to the EU AI Act’s disclosure requirements for AI-generated content. The mechanism nudges word choice among plausible alternatives rather than inserting anything visible, which is reportedly what lets it survive copy-paste and light editing.

A detection API exists for verifying watermarked content, but it isn’t public — access is limited to regulators, law enforcement, media organizations, fact-checkers, and vetted independent researchers. For most developers this will be invisible in practice; it’s worth knowing about mainly if you’re shipping content into a regulated or EU-facing context.

Why this matters if you’re shipping AI features

If you’re running an agent loop, a coding assistant, or anything else that leans on repeated context via prompt caching, this is a direct cost reduction with no code changes required — it’s worth re-checking your margins on any AI feature priced before September 1. The watermarking change is lower-stakes for most teams, but if you operate in the EU or serve regulated industries, it’s a compliance detail worth flagging to whoever owns that risk.

Sources

• gHacks Tech News — Anthropic Releases Claude Fable 5.1 and Cuts Cached Token Pricing by 75%

• TechSpot — Anthropic releases Claude Fable 5.1, cuts cached token pricing by 75%

• MacRumors — Anthropic Launches Claude Fable 5.1 With Lower Costs and Fewer False Positives

• TNW — Anthropic releases Claude Fable 5.1 and Mythos 5.1, cutting cache read prices by 75%