Claude Fable 5.1 and Mythos 5.1: specs, benchmarks, and what changes for us

On 1 September 2026 Anthropic released two models: Claude Fable 5.1 for all customers and Claude Mythos 5.1 for Project Glasswing participants only. What follows is the full specification, every published result, the comparison against the rest of the family, and the pricing. At the end, the answer to the question we care about most: whether any of it changes how we work.

Two models, one specification

Fable 5.1 replaces Fable 5 and Mythos 5.1 replaces Mythos 5. Both new models carry identical specifications and identical prices. Only eligibility separates them.

Per Anthropic documentation, accessed 2 September 2026.

Fable 5.1 runs on the Claude API, Amazon Bedrock (as anthropic.claude-fable-5-1), Google Cloud, Microsoft Foundry and Claude Platform on AWS. Both models require 30-day data retention and are unavailable under zero data retention without separate authorization from Anthropic.

Benchmarks Anthropic published

Every figure below comes from the vendor announcement, so read them as vendor benchmarks. We have no independent measurements of these models and will not pretend otherwise.

Per the Anthropic announcement of 1 September 2026, accessed 2 September 2026.

The biggest jump is scientific research, more than doubling its predecessor and clearing Opus 5 by twenty-four points. AutomationBench, covering business workflows, comes second and nearly doubles too, though an absolute score of 31.4 percent shows how far these tests remain from saturation. Reasoning gains are modest by comparison, running to about three points, and on CursorBench the margin over Opus 5 is 3.4 points at twice the price. The shape of these results says more than the peaks: the model pulls clearly ahead where a task runs long and takes many steps, and draws roughly level where one good answer is the whole job.

One detail in the data deserves noting. The documentation says Mythos 5.1 has the same capabilities as Fable 5.1, yet the announcement lists Mythos separately on one benchmark, Terminal-Bench 4.0, at 60.9 percent against 55.8 for Fable. Anthropic does not explain the gap, so neither will we.

How it sits against the rest of the Claude family

Fable 5.1 is not the default choice, and Anthropic does not present it as one. Their guidance is to start with Opus 5 and move up for demanding reasoning and long-horizon agentic work.

Per the models overview in Anthropic documentation, accessed 2 September 2026.

Fable 5.1 costs twice what Opus 5 does, on an identical context window and an identical output limit. You are paying for reasoning quality and a fresher knowledge cutoff, not for capacity. The arithmetic is plain enough: a million input tokens plus a hundred thousand of output runs to $15 on Opus 5 and $30 on Fable 5.1. On a single job that difference is noise; across a thousand runs a month it becomes a budget line somebody notices. The fresher cutoff, June rather than May 2026, is one month of difference and rarely decides anything on its own.

Pricing and the promised savings

Input and output cost exactly what they did on Fable 5. One line moved.

The cache read multiplier is now 0.025 of the base input price, against 0.1 on every other Claude model. Anthropic estimates that this makes Fable 5.1 run roughly 25 percent cheaper than Fable 5 on typical workloads, and up to 45 percent on heavily agentic ones.

Those figures compare against Fable 5, not against Opus 5. Measured against Opus you still pay double on both input and output.

API changes

Three things stop working and five arrive. The order matters only to teams already calling Fable 5.

• Forced tool use is gone: tool_choice of type any or tool returns a 400. Auto and none remain. The reason: thinking is always on, and a forced call would skip it.
• Thinking blocks belong to their model: Fable 5.1 reads earlier models’ blocks, none of them read its own. Switching models mid-conversation drops the blocks silently and unbilled.
• Editing history invalidates thinking: changing the system prompt, the tool list or an earlier message raises an error. Enforced for accounts created on or after 31 August 2026.

Arriving alongside them: changing effort mid-conversation without invalidating the cache, system messages scoped to a single turn, readable progress updates between tool calls, the cheaper cache reads, and content provenance marking. The first three are in beta and need their own headers.

The watermark deserves its own note. Every piece of text from these models carries Anthropic’s statistical watermark, and image and video files carry signed C2PA credentials. The vendor states it adds no tokens or hidden characters and does not alter content. How search engines will treat it is something nobody knows today.

Seven behaviours that changed on their own

This is the most interesting list in the documentation, because it covers things that shift with no code change at all. Swapping the model ID passes green and the output differs.

• Less parallel tool calling: one tool per turn where Fable 5 batched several. More turns and tokens, unchanged quality.
• Fewer progress updates: less narration of what it is doing, especially at high effort.
• Less retrieval at low effort: it answers from memory rather than searching.
• Denser prose: longer sentences and fewer paragraph breaks.
• Less formatting: bold, headings and lists used more sparingly.
• Unmarked quotations: when summarizing, it repeats source passages without marking them as quotes.
• Whole-file rewrites: a full rewrite instead of a targeted edit, which raises the output bill.

Science results from Project Glasswing

Anthropic reports three results from the closed research programme. None can be independently verified, so treat them as vendor claims.

• Protein design: binding affinities ten times higher than the best competition designs, at a hit rate near 50 percent against a typical 10 to 15.
• Venus mapping: detail down to 2 or 3 kilometres rather than 10 to 20, with heights up to 25 percent more accurate.
• Computational biology: model speedups up to 2.5 times, with estimated GPU cost savings of 30 to 60 percent.

What changes at Seedlight

Here we stop quoting the vendor. We checked where those three breaking changes could reach us, and the answer is short: not one of them applies. There is nothing for us to migrate, and a release on its own is no reason to change that.

In how we build: possibly

The second place these models touch us is Claude Code, which we build platforms with every day. There the harness keeps the conversation history intact for us, so the third breaking change is handled on our behalf.

Two things remain worth watching. The first is cost: on long agentic sessions the cheaper cache read genuinely cuts the bill, because an agent working through a catalog rereads the same prefix on every pass. The second is quality on long tasks, which is precisely where the benchmarks show the largest gains.

The counterweight sits in the same documentation, and it is worth doing the sums before anyone announces a saving. Cache reads fall by 75 percent, but they fall on a line that costs $0.25 per million. Rewriting whole files instead of making targeted edits raises the output bill, and output runs at $50 per million, two hundred times dearer. It takes only a little more rewriting than the reading saved for the entire discount to vanish. Those two effects pull against each other, and I am not going to claim I know which wins for us before measuring it.

What we are not doing

We are not moving anything to Fable 5.1 because a release happened. It costs twice what Opus 5 does, and its advantage sits in areas our work does not: scientific research and very long agentic runs. On an eCommerce platform build, that gap rarely returns the doubled bill.

If you run AI automation of your own, the same arithmetic is worth doing before you swap a model ID. We described the pattern of silent behavioural drift in our piece on why AI automations fail quietly, and how we structure agent work in our piece on agentic engineering.

At Seedlight the model choice lives in the AI Automation stage of the BEAM framework, alongside evals on the client’s own data. That makes changing model a one-line change and a test run. It does not remove the behavioural risks above; it only shortens the time before you find out about them. The scope is on our AI commerce automation page, and for generated copy specifically, product descriptions.