Claude Fable 5.1 is out - a 75% cache-read price cut aimed at always-on agents

What happened

Anthropic released Claude Fable 5.1 on September 1. It is the company's top model for coding and agent work, and the headline is not a benchmark. It is the price of cache reads, which dropped from $1.00 to $0.25 per million tokens, a 75% cut. Base pricing is unchanged: $10 per million input tokens, $50 per million output, a 1 million-token context window, 128K maximum output. There is a second model, Mythos 5.1, which Anthropic describes as the same model with looser safeguards, available only to vetted cybersecurity and life-sciences organizations in the US.

According to Anthropic's announcement, Fable 5.1 scores 52.6% on Terminal-Bench-Science (up from Fable 5's 24.7%) and 55.8% on Terminal-Bench 4.0 (up from 42.0%). VentureBeat reported early-access results: Ramp ran an unattended 38-hour machine-learning job across six experiments, and a Millennium team used the model to trace a rare software crash to a bug in a vendor library.

What is genuinely new

The cache-read cut is the part that changes budgets. When an agent works on a task for hours, it re-reads the same context on every step: the system prompt, the tool definitions, the codebase or document set it was handed. That context is cached. Before, you paid $1.00 per million tokens each time it was read back. Now you pay a quarter of that. Anthropic estimates the effect at about 25% lower cost for a typical workload and up to 45% for heavily agentic work, where the same context is read hundreds of times.

The benchmark gains matter less for most businesses. A jump on a science-terminal test tells you the model got better at a narrow, hard problem. It does not tell you it will handle your invoice exceptions or your support triage. Treat those numbers as a sign the model improved, not a promise about your workflow.

What it means for a business owner

If you run an agent that does one thing quickly, like classifying an email or pulling a field from a PDF, the cache discount barely touches you. Your context is small and read once.

The savings land on long-running agents. A pipeline that reviews a full codebase before making a change. An agent that works through a 200-row spreadsheet, keeping the whole instruction set and schema in context on every row. A research task that runs for an hour. These are the jobs where context is large and re-read constantly, and where a 45% cost drop can move a project from "too expensive to run daily" to "run it every morning."

It also makes unattended runs more realistic to budget for. A 38-hour job that used to cost several hundred dollars in cache reads alone now costs a fraction of that. That does not make it safe to run unattended. It makes it cheap enough to consider, which is a separate question you still have to answer.

The honest caveat

Fable 5.1 is still a premium model. At $10 input and $50 output, it is more than twice the promotional rate of OpenAI's GPT-5.6 Sol and well over ten times Gemini 3.7 Flash on output, according to pricing VentureBeat compiled. The cache discount narrows the gap for agent work. It does not close it. If your task is simple and high-volume, a cheaper model is still the right call, and Fable 5.1's benchmark lead will not show up in your results.

The cache savings are also conditional. You only get them if your pipeline is built to reuse cached context: same prompt prefix, same tool set, cache kept warm between calls. A pipeline that rebuilds its context from scratch every call pays the discount on nothing. Check how your integration is structured before you assume the 45% applies to you.

Mythos 5.1 is gated. If your use case is defensive security work, you have to apply and be verified, and only US organizations qualify right now.

What to do about it

Look at your single most expensive agent, the one with the biggest monthly bill. Find out what share of its token spend goes to cache reads. If it is a long-running or context-heavy job, that number is probably large, and moving it to Fable 5.1 is worth testing this week. If it is a short, simple task, leave it where it is and spend the attention somewhere it will pay off.