Claude Fable 5.1 and Mythos 5.1: What Changed
Anthropic shipped Fable 5.1 and Mythos 5.1 on 1 September: one model, two safeguard tiers, and a cache read price that fell from $1.00 to $0.25 per million tokens.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. The headline number for anyone running agents is not a benchmark: it is the cache read price, which fell from $1.00 to $0.25 per million tokens. Input stays at $10 and output at $50 per million. Anthropic puts the saving at roughly 25% for typical workloads and up to 45% for highly agentic ones. The second surprise is structural. Fable 5.1 and Mythos 5.1 are the same model shipped under two names with different safeguard settings.
The pricing change matters more than the benchmark table
Cached input is the part of your prompt the provider has already processed and can replay cheaply. In a long agent loop, it is most of your bill. An agent that reads a 40,000 token codebase context on every one of 30 steps pays for that context 30 times, and after the first pass almost all of it is a cache hit.
Work the arithmetic. Thirty steps, 40,000 tokens of stable context, one cold read and 29 cached ones:
Line item | At Fable 5 cache price | At Fable 5.1 cache price |
|---|---|---|
Cold read (40k at $10/M) | $0.40 | $0.40 |
29 cached reads (1.16M at cache rate) | $1.16 | $0.29 |
Context subtotal | $1.56 | $0.69 |
That is a 56% cut on the context portion of one agent run, before you count the fresh tokens each step adds. It is also why the saving is bigger for agentic work than for chat: chat sessions are short and mostly fresh input, agent loops are long and mostly repeated input. If you have never measured your cache hit rate, that number now decides more of your bill than your model choice does. The mechanics are covered in our explainer on what prompt caching is and when it pays.
One model, two names, two safeguard tiers
Fable 5.1 is generally available on the Claude API, AWS, Google Cloud and Microsoft Azure. Mythos 5.1 is the same weights behind trusted access programs, currently for US organisations, through a Cyber Verification Program for defensive security work and a Life Sciences Verification Program for research professionals.
This is worth naming clearly because it is becoming an industry pattern rather than an Anthropic quirk. Google did the same thing on 2 September with Gemini 3.8 Flash and a separately gated Flash Cyber variant. The capability is one thing, the permission to use it in sensitive domains is another, and vendors are starting to sell them as separate products. If your product needs the permissive tier, you are now applying for access rather than paying for a higher plan. Anthropic previously used the same split for Fable 5's biology safeguards.
On the security side, Anthropic says the new safeguards block 60% fewer false positives than before, and that Fable 5.1 can be used to discover software vulnerabilities but not to develop exploits for them. Under its Responsible Scaling Policy the model falls short of the next risk tier, so it ships with the same safeguards as Mythos 5.
The benchmarks, and what they actually tell you
Anthropic's published comparison against Fable 5:
Benchmark | Fable 5.1 | Fable 5 |
|---|---|---|
Terminal-Bench-Science 0.1 | 52.6% | 24.7% |
Terminal-Bench 4.0 | 55.8% | 42.0% |
Humanity's Last Exam (no tools) | 60.9% | 57.8% |
CursorBench 3.2.0 | 73.4% | 70.5% |
OSWorld 2.0 (strict) | 41.7% | 36.1% |
Note the shape of the gains. The two Terminal-Bench numbers, which measure long-running work in a real shell, moved by 13.8 and 27.9 points. The knowledge benchmark moved 3.1. That is a release aimed at agents that run for a while, not at models that answer questions. Terminal-Bench-Science more than doubling is the single loudest number in the table, and it is the one closest to what a coding agent actually does all day.
The usual caveat holds: vendor-published benchmarks are chosen by the vendor. The way to know whether this helps you is to test the new model against your own workload before switching, not to read the table twice.
What to do this week
Check your cache hit rate before you change anything else. If it is low, the price cut does nothing for you and fixing your prompt structure is worth more than the upgrade.
Re-run your own evals rather than trusting the delta. A 27-point jump on Terminal-Bench does not predict a 27-point jump on your test suite.
If you are in defensive security or life sciences, look at whether the trusted access route is worth the paperwork. If not, Fable 5.1 does the same work with tighter refusals.
Re-do your cost model. A 25% to 45% swing changes what is worth automating, and it is the kind of change that quietly makes a previously unprofitable feature profitable.
For the wider pattern of how to read a release like this without getting swept along, see how to read an AI release note without the hype and our running guide to keeping up with AI news. If you are comparing this against other providers, the API pricing comparison guide covers the traps in per-token maths.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
