Claude Fable 5.1 and Mythos 5.1: what changed, and why the cache price matters more than the benchmarks

Claude Fable 5.1 and Claude Mythos 5.1 launched on September 1, 2026, three months after Fable 5 arrived in June as the first Mythos-class model the public could use. Fable 5.1 is available to paid Claude users and every API account. Mythos 5.1 is the same underlying model with looser safety filters, limited to a set of US organisations, with no timeline for wider access. Anthropic's own description of the split is the cleanest thing about the release: Fable declines requests that touch cyberattacks or dangerous biology, and Mythos exists for the organisations that need those capabilities and can be trusted with them.

The benchmarks that look like work

Anthropic has a habit of publishing benchmarks that resemble actual jobs, and the 5.1 figures from Tech-ish's launch coverage are unusually easy to read. Terminal-Bench-Science, a test of research tasks run from a terminal, more than doubled. Terminal-Bench 4.0, a coding test, rose fourteen points, with Mythos 5.1 five points higher again. AutomationBench, which measures business workflows end to end, nearly doubled. Knowledge now runs to June 2026, five months later than Fable 5.

AutomationBench is the row a business owner should stare at. It measures whether a model can carry a multi-step business workflow through to a finished result, and a jump from 17.1 to 31.4 percent is the difference between an agent that needs a babysitter and one that needs a reviewer. It is still 31.4, not 90, which is why the approval gate stays in every automation Monolith installs. But the direction and the size of the move are real.

The number that changes the bill

List prices did not move: $10 per million input tokens and $50 per million output. What moved is the price of a cache read, from $1.00 to $0.25 per million tokens. Anthropic's arithmetic is that most customers end up around 25 percent cheaper, and cache-heavy workloads save up to 45 percent. Cache reads are what you pay when the model re-reads the same system prompt, brand guide, knowledge base, or codebase on every call. For an agent that runs a workflow forty times a day against the same client context, that is nearly the whole bill.

Context stays at one million tokens and output at 128,000. The practical consequence: it is now cheap to give an agent the whole brand, the whole process doc, and the whole client history on every run, provided the stable material is arranged so it caches. That is a prompt-architecture job, and it is one of the first things Monolith rewires when it takes over an existing automation.

Watermarks, anti-distillation, and a filter that fires less

• Every output now carries an invisible statistical watermark, Anthropic's compliance move for the EU AI Act's rules on models released after August 2, 2026. It contains no user data.
• New API accounts can no longer edit Claude's past messages while keeping its underlying reasoning, an anti-distillation restriction aimed at competitors training on Claude's outputs.
• The cybersecurity filter fires about 60 percent less often per session and the biology filter about 85 percent less on ordinary medical questions, while still refusing exploit generation. Three external firms, including Gray Swan, attacked the new filters and reported no critical jailbreak.
• A zero-retention option is available for enterprise data, with a further set of enterprise safeguards promised for later in 2026.

The filter change matters more for day-to-day work than it sounds. A model that refuses a routine question about a medication interaction or a firewall rule is a model your team stops trusting for the questions it should answer. Fewer false refusals means fewer workarounds, and fewer workarounds means fewer people pasting company data into whichever tool did not say no.

What Monolith is doing with it this week

Fable 5.1 is the model behind Claude Code, which is one of the two agent harnesses Monolith runs daily alongside Hermes. The jump on terminal and automation benchmarks shows up directly in the work: longer unattended runs, fewer restarts, more of a workflow finished before a human looks at it. The cache price shows up in the invoices. And the Mythos line is a useful thing to explain to clients who ask whether they are getting the real model: yes, Fable 5.1 is the Mythos-class model, and the difference is a filter set, not a smaller brain.