Claude Fable 5.1 and Mythos 5.1: What Shipped September 1

Analysis

What shipped

Claude Fable 5.1 and Claude Mythos 5.1 launched on September 1, 2026. Fable 5.1 keeps the Fable price point at $10 input and $50 output per million tokens, a 1M token context window, 128K max output, and always-on adaptive thinking with a default effort of high at the platform level. Its knowledge cutoff is June 2026 and its published retirement floor is no sooner than September 1, 2027.

Mythos 5.1 is the same underlying model with more permissive safeguards, available only through trusted access under Project Glasswing, at the same $10 / $50 rate with the same $0.25 cache reads. Anthropic also began enrolling the first participants in Mythos life-sciences trusted access as part of the scientists program.

Analysis

The cache-read cut is the real pricing story

Base rates did not move, so a surface reading of the announcement suggests nothing changed on cost. That is wrong for agentic workloads. Cache reads on Fable 5.1 and Mythos 5.1 bill at 2.5% of base input — $0.25 per million tokens — against the 10% that applies on most other models. Anthropic quantifies the effect as roughly 25% savings on typical workloads and up to about 45% on highly agentic ones.

The shape of the saving matters: it accrues to workloads with a large, stable cached prefix that gets re-read many times. Long-running agents, repository-scale coding sessions, and document-heavy analysis loops are the profile. A single-turn chat request sees almost none of it. Before you dismiss Fable 5.1 as the expensive option, compute your own cached-read share.

• Cache read: 2.5% of base input on Fable 5.1 and Mythos 5.1, versus 10% on most models.
• Base rate: unchanged at $10 input and $50 output per million tokens.
• Batch API: 50% discount still available for non-interactive work.
• Measure your cached-read-to-total-input ratio before estimating the savings.

Analysis

Breaking changes and additive betas

The platform overview documents three breaking behaviors relative to Fable 5. Forced tool use now errors. Earlier models cannot read Fable 5.1 thinking blocks. Editing an earlier turn invalidates thinking. Each of these can surface as an intermittent production failure rather than a clean error at deploy time, which makes them worth testing explicitly.

The additive changes are opt-in betas: per-message effort, turn-scoped system messages, a display: "updates" progress mode, the lower cache-read rate, and content provenance. Per-message effort is the one most likely to change architecture, because it lets a single conversation mix cheap and expensive turns instead of committing to one effort level.

• Test forced tool use paths before migrating; they now error rather than degrade.
• Audit any multi-model pipeline that passes thinking blocks between models.
• Check flows that let users edit earlier turns, including retry and branch features.
• Consider per-message effort before splitting a workload across two models.

Analysis

Data handling: EFS and zero data retention

Anthropic paired the launch with Enterprise Frontier Safeguards, a customer-controlled storage arrangement being phased in through the fall. Until EFS is available to an account, eligible customers can use Fable 5.1 under zero-data-retention terms. For regulated buyers this is the gating detail, not the benchmark table: it determines whether the flagship can touch the workload at all.

If your procurement path depends on either EFS or ZDR, confirm current eligibility with Anthropic directly rather than inferring it from the announcement. The phased rollout means account-level answers differ from the published description.

Analysis

Anthropic-reported benchmarks

The figures below are from Anthropic’s launch announcement. Claude Weekly reproduces them as published claims, not as independent measurement, and does not analyze them further — benchmark analysis is out of lane for this site.

Analysis

Fable 5 history, in its place

Fable 5 and Mythos 5 launched on June 9, 2026, access was suspended, and Anthropic published a redeployment note on June 30 and a safeguards follow-up on July 2. That sequence taught a useful lesson — flagship availability can be a policy event rather than an infrastructure event — and it is now history. Fable 5 is a legacy model on the platform listing.

Keep the operational habit and drop the framing. When flagship access changes, check Claude Status, app release notes, and the newsroom before assuming a client bug. But do not describe the June suspension as the current Fable story.

Analysis

Safety routing you should know about

Biology-related blocks on Fable now route to Opus 5 rather than Opus 4.8, and Opus 5 introduced automatic fallbacks when safety classifiers flag Opus 5 or Fable 5. Anthropic also states that Fable continues to block professional biology and drug-development work, with biology and chemistry researchers in the scientists program limited to Opus-class models.

The published cyber policy is that classifiers permit finding vulnerabilities in source code while blocking binary-based scanning, penetration testing, and exploit generation. Claude Weekly reports the policy boundary; it does not publish guidance for the blocked categories.