Claude Fable 5.1 and Mythos 5.1: What's New, Explained

On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 — the same underlying model shipped in two safeguard tiers. Fable 5.1 is generally available today; Mythos 5.1 is the same model with lighter guardrails, restricted to vetted cybersecurity and life-sciences organizations. If you only remember one thing: it's smarter than Fable 5 at long agentic tasks, meaningfully cheaper for context-heavy work, and its research results are the headline.

Here's what actually changed and whether it matters for how you build with agents.

Fable 5.1 and Mythos 5.1: one model, two tiers

Anthropic ships the pair as a single model behind two sets of safeguards, per its announcement:

• Fable 5.1 — the general-availability version with production safeguards. This is what you get on the API (claude-fable-5-1), on Claude.ai, in Claude Code and Claude Cowork, and via AWS, Google Cloud, and Azure.
• Mythos 5.1 — the same weights with more permissive guardrails, available only through two trusted-access programs: a Cyber Verification Program for defensive security work and a Life Sciences Verification Program run in partnership with the US government.

The naming split matters because the two "models" score differently on some benchmarks purely because Fable's safeguards intervene on certain tasks. Anthropic says its safeguard improvements should shrink that gap going forward.

In Anthropic's lineup, Fable sits above Haiku, Sonnet, and Opus as the most capable tier.

What's new vs Fable 5

Three things stand out over the June 2026 Fable 5 release: cost, agentic strength, and science.

1. Cheaper, mostly through cache reads

Fable 5.1's headline price cut comes from cache reads — where the model reuses context it already processed. Anthropic dropped that price by 75%, to $0.25 per million tokens. The knock-on effect:

• ~25% lower cost for typical workloads billed by token.
• Up to ~45% lower for highly agentic, context-heavy work where cache reads dominate the bill.

Everything else is unchanged: $10 per million input tokens and $50 per million output tokens. So if you run long-horizon agents that re-read large contexts, this is a real cut; if your calls are short and output-heavy, less so.

Fable 5.1 also matches or beats Fable 5's quality at Low or Medium effort — meaning you can often dial effort down and pay less for the same result. (Note the defaults differ: High in Claude Code, Medium in Claude Cowork and on Claude.ai.)

2. Better at long, agentic work

The clearest jump is on long-running, tool-heavy tasks. On Anthropic's published benchmark table, Fable 5.1 improves over Fable 5 across coding, computer use, and business workflows:

Source: Anthropic. The biggest single gap is agentic scientific research, where the score more than doubles over Fable 5.

Benchmarks aside, the qualitative claim is that Fable 5.1 stays readable and on-task over long runs and fixes root causes instead of taking shortcuts. Anthropic cites the investment firm Millennium finding Fable 5.1 traced a one-in-a-million crash — by disassembling a vendor library and matching it against a core dump — that had gone unexplained for years and that no other model, including Fable 5, could crack.

3. Science is the real story

The part worth paying attention to is the research work, done largely on the Mythos 5.1 tier:

• Protein design: given open-source folding and design tools, Mythos 5.1 designed high-affinity binders with a hit rate near 50% across 12 targets — Anthropic says typical hit rates today are 10–15% — and beat the best entries in Adaptyv Bio's competitions on three targets by roughly 10x affinity.
• Planetary mapping: Fable 5.1 trained a neural network to build a higher-resolution elevation map of a third of Venus from 30-year-old NASA Magellan radar, sharpening detail from 10–20 km down to 2–3 km. Anthropic is releasing it under a Creative Commons license.
• GPU kernels: Mythos 5.1 wrote custom kernels that sped up seven open-source genomics and protein models by up to 2.5x with identical outputs, cutting estimated GPU costs on genome-wide analyses by 30–60%.

These are demos, not peer-reviewed results — but they're concrete, externally validated in places, and point at where agentic models are heading.

Cost, data retention, and safeguards

Beyond raw capability, Anthropic tied the release to three long-standing enterprise complaints.

Data retention. A new system called Enterprise Frontier Safeguards (EFS) gives customers zero-data-retention-level privacy by storing data on the customer's own cloud infrastructure, with human review done by the customer. It rolls out in phases starting this fall; until then, eligible customers can run Fable 5.1 with zero data retention.

Fewer false positives. Safeguards now flag benign content less often — Anthropic says cybersecurity safeguards block 60% fewer false positives, and Claude Code users should see roughly 60% fewer cyber-safeguard interventions per session. Fable 5.1 can now be used to find software vulnerabilities (defensive work), though exploit generation, penetration testing, and binary vulnerability scanning still get routed to Opus models.

Anti-distillation. New API accounts can no longer edit Claude's prior context while preserving its thinking transcript — closing a documented technique used to extract reasoning traces. Existing accounts aren't affected yet, but the change applies to all users on future releases.

Should you switch?

A quick read by use case:

• Long-horizon coding agents — likely yes. The agentic gains plus cache-read savings are exactly aimed at multi-step, context-heavy runs. Cognition (Devin) said it's moving Opus 5 traffic to Fable 5.1 at launch, starting with code review.
• Short, output-heavy calls — the price cut is smaller here since it's driven by cache reads, not output tokens. Test before assuming savings.
• Cybersecurity or life-sciences R&D — you'll hit Fable's routing to Opus for dual-use tasks. Mythos 5.1 exists for exactly this, but only via the vetted-access programs.
• Privacy-sensitive enterprises — EFS is the headline for you, but it's phased in over the fall, so confirm availability on your platform before committing.

For most builders, the honest summary is: same top-tier quality, cheaper on the workloads that used to be expensive, with guardrails that get in the way less often.

Do this with your own AI workforce

Fable 5.1's real edge is long-running, tool-heavy work — the same shape as the multi-agent tasks that fill a real workday. Eigent is an open-source "Cowork" desktop app that runs a multi-agent AI workforce locally, so you can point frontier models like Claude at coding, research, and document workflows on your own machine — a natural fit if data retention is a concern. See how it handles multi-step engineering in our review GitHub PRs workflow, or download Eigent and wire up your own agents today.