Contents
On September 1, Anthropic shipped Claude Fable 5.1 alongside a restricted twin called Mythos 5.1. The headline number is a 75% price cut on prompt cache reads, but the more interesting story for anyone running agents in production is what that price cut is actually rewarding: workloads that keep a model’s context warm across many turns instead of re-sending it from scratch.
This matters because prompt caching has quietly become the load-bearing cost lever for agentic systems. An agent that reads the same tool definitions, system prompt, and file context on every step pays for that context repeatedly unless it hits cache. Anthropic just made hitting cache four times cheaper, which changes the economics of exactly the workloads, long-running coding agents, multi-step research loops, terminal automation, that were the most expensive to run at scale.
What actually changed
Cache reads dropped from $1.00 to $0.25 per million tokens. Base input and output pricing stayed put at $10 and $50 per million tokens, the same list price as Fable 5. Five-minute cache writes are $12.50 per million tokens and one-hour cache writes are $20. Anthropic estimates the net effect at roughly 25% cheaper for typical workloads billed by token, and up to about 45% for heavily agentic ones that keep a large fraction of their input coming from cache.
Terminal-Bench-Science more than doubled, from 24.7 to 52.6. On Anthropic’s own launch benchmarks, Fable 5.1 leads both Fable 5 and Opus 5 across the board, and leads GPT-5.6 Sol on every benchmark where a direct comparison is available. Terminal-Bench-Science specifically measures multi-step scientific and technical tasks executed through a terminal, the kind of long tool-use chain that benefits most from cheaper caching, so the pricing and capability changes are pointed at the same use case rather than being two unrelated announcements bundled together.
Fable and Mythos are one model with two safety configurations, not two models. Fable 5.1 runs with Anthropic’s full consumer-facing safety classifier stack and is generally available on the API, Claude.ai, Claude Code, and Claude Cowork, and through AWS, Google Cloud, and Microsoft Azure. Mythos 5.1 runs the same underlying weights with those classifiers relaxed specifically for cybersecurity and life-sciences work, and it ships only to organizations Anthropic has separately vetted through a trusted access program. Anthropic rates the shared model at CB-1 for chemical and biological capability and has moved its overall alignment-risk assessment from “very low” to “low,” which is worth noting if you are deciding whether your team needs Mythos access at all versus just using Fable.
Try it: what the cache cut is worth to your workload
About the calculator
This uses Anthropic’s published list prices for Fable 5 and Fable 5.1. “Cache share” is the percentage of your input tokens served from a warm cache rather than sent fresh, which is the number that determines how much of the 75% cut you actually collect.Move the slider to see blended cost per million input tokens.
Drag the slider past 85% and the savings curve lands right in Anthropic’s own “up to 45%” claim for heavily agentic workloads. Below about 30% cache share, the cut barely moves your bill, which is the real qualifier behind the headline number: this release rewards architecture, not just an upgrade.
What this means for you
If you are running an agent loop that re-sends a large system prompt, tool schema, or codebase context on every step, check whether you are actually hitting Anthropic’s prompt cache before assuming the price cut helps you. A model swap to Fable 5.1 with no caching strategy nets you nothing on the input side, since base rates did not move. My own next step is auditing the cache-hit rate on our longer-running Claude Code sessions before deciding whether Mythos-class access is worth pursuing for anything security-adjacent we run, versus just restructuring prompts so Fable 5.1’s cache pricing actually pays off.