Claude Fable 5.1 is out: cheaper in practice, same price on paper

Anthropic released a new version of its most capable model on 1 September 2026, and did something slightly unusual with the pricing: it left the advertised numbers exactly as they were, then cut a different cost that most people never look at but almost everyone pays. The net effect for most users is a bill roughly a quarter smaller.

When you have a long conversation with an AI, or when an AI tool works through a big task, the same background information gets sent to the model again and again — the instructions, the documents, everything said so far. Providers store that repeated material in a cache so it does not have to be processed from scratch each time, and they charge a reduced rate for reading from it.

That cache rate is where a surprising share of real bills come from. Anthropic cut it by three quarters.

In practice, Anthropic estimates that works out to about 25% less than the previous version for typical use, and as much as 45% less for tasks where an AI works through something long and complicated — writing and testing software, for example, or researching across many documents.

The new version also lets you choose how hard it thinks. At low or medium effort, Anthropic says it performs as well as or better than the previous model while costing considerably less. At high effort it goes further than the old model could. Different products pick sensible defaults: high effort in Claude Code, where the tasks are demanding, and medium in the general chat interface.

Anthropic released two models with different names that are, underneath, the same thing. Fable 5.1 is the version everyone can use. Mythos 5.1 is the identical model with fewer restrictions, available only to organisations that have been vetted — mainly in cybersecurity and life sciences, where legitimate work sometimes runs into safety filters designed to stop misuse.

For everyone else, that means the standard version sometimes declines requests touching cybersecurity, biology and chemistry, or copying other AI models — and those requests get handled by a less capable model instead. It is a sensible compromise, and it is becoming the industry norm: OpenAI runs a comparable vetted-access programme for security work.

One more thing worth knowing if you compare models: Anthropic says these safety measures can affect benchmark scores, so published test results may not exactly match what a normal user experiences. That is an honest disclosure and a good reminder that benchmark tables deserve a pinch of salt regardless of who publishes them.

Among the performance claims, one stands out because it can be checked against real human work. Anthropic says that on three targets, the model designed proteins that bound to their targets ten times more strongly than the best entries submitted to a competition run by a company called Adaptyv Bio.

Protein design matters because it underpins drug development: if you can design a molecule that attaches tightly and specifically to a target in the body, you have the starting point for a medicine. A result ten times better than the best human-submitted entries is a meaningful signal that these systems are becoming genuinely useful in laboratory science, not just in writing and coding.

It is also a reminder of why the restricted Mythos version exists. The same capability that designs a useful protein is capability worth being careful with, which is why access to the less-restricted version is limited to vetted organisations in fields like life sciences and cybersecurity. The trade-off is real and the industry is, for now, handling it by checking who is asking.