Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, and the two are the same model. The only thing separating them is how tightly the safety layer is wound: Fable 5.1 ships to everyone with cybersecurity and life-sciences restrictions active, while Mythos 5.1 loosens those restrictions for organizations Anthropic has vetted individually. Alongside the release, Anthropic cut the price of cache reads by 75 percent and opened software vulnerability discovery to the public model for the first time.
Key facts
- Cache reads dropped 75 percent, to $0.25 per million tokens; base input ($10) and output ($50) per million tokens are unchanged.
- Claude Code users should see roughly 60 percent fewer interventions per session from Anthropic's cyber safeguards, according to the company.
- Announced September 1, 2026; the Hacker News thread drew 968 points and more than 900 comments the same day.
- Primary source: Anthropic's launch post and the Fable 5.1 model documentation.
The two-names-one-model structure is the interesting part, and it is not new -- Anthropic did the same thing with Mythos 5 earlier this year. What is new is how explicit the company has become about it. "Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards," the launch post says. Read plainly, that is Anthropic conceding that the thing it sells is not really a model at all. It is a model plus a policy, and the policy is the product line.
To understand why that matters, it helps to know what a safeguard actually is here. It is not a change to the neural network. It is a separate classifier watching the conversation, and when it decides a request looks like weapons research or offensive hacking, it either refuses or quietly hands the task to a different, less capable model. Think of it as a bouncer standing outside a room. The person in the room is the same either way; what changes is who the bouncer lets through the door. Mythos 5.1 is the same room with a more permissive bouncer, and you have to apply to get on the list.
That bouncer has been too aggressive, and Anthropic is admitting it. Security researchers using Claude to audit their own code kept getting stopped. The company says its updated cyber safeguards now interrupt Claude Code sessions about 60 percent less often, and that its biology safeguards "fire 85% less often for benign requests related to elementary biology and medical questions." Fable 5.1 is now allowed to identify vulnerabilities in source code -- the defensive half of security work. Exploit development, penetration testing, and binary vulnerability scanning still get routed away to Opus models.
The performance claims come with an unusual footnote that is worth pausing on. Anthropic says Fable 5.1 was benchmarked with its production safeguards switched on, and that on tasks where the guardrail fired, the model scored zero or the work was handed to an older Opus model. The published numbers are therefore not a ceiling. They describe the model as customers will actually experience it, guardrails and all -- a more honest framing than most benchmark tables get, and one that quietly makes the scores harder to compare against competitors who publish unrestricted numbers. If you want the background on why that distinction matters, our explainer on how AI systems get benchmarked covers it.
The pricing story deserves care, because the headline is doing work the numbers do not support. Anthropic says Fable 5.1 costs "an estimated 25% less than Fable 5 for typical workloads" and up to 45 percent less for agentic work. But it did not cut the sticker price. Input stays at $10 per million tokens, output at $50. The whole discount lives in cache reads -- the cheap re-reading of context the model has already processed, explained in our lesson on prompt caching -- which fell to $0.25 per million. The independent benchmarking firm Artificial Analysis then measured what a finished task actually costs and got the opposite answer: $3.69 per task at maximum effort, against $3.14 for Fable 5. The reason is that Fable 5.1 talks more -- roughly 1.7 times the output tokens at max effort -- and output tokens are the expensive kind. Both things are true. Cheaper per cached token, pricier per hard task finished.
Simon Willison, testing on launch day, found the same lever from the user side: the model's behavior swings sharply with the effort setting, skipping reasoning on simple tasks at low effort and spending heavily at max. The dial that matters is effort, not the version number.
Early partners were enthusiastic in the specific way that reads as real. Craig Falls, Head of Quantitative Research at Jane Street Capital, said that "while prior models became hard to follow the longer they worked, Fable 5.1 remains readable over long, multi-step tasks." A senior portfolio manager at Millennium described the model disassembling a vendor library, matching it against a core dump, and finding a one-in-a-million crash nobody had explained in four or five years.
The honest caveat: every number above except the Artificial Analysis measurement comes from Anthropic or from partners Anthropic selected and quoted. Twenty-two testimonials on a launch page are marketing, however credible each individual account sounds. The independent cost measurement already contradicts the company's own framing on the axis customers care about most, which is a reasonable prompt to wait for third-party evaluations before believing the rest.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)