Claude Fable 5.1 and Mythos 5.1 for coding agents

Anthropic shipped two frontier models on September 1: Claude Fable 5.1 and Claude Mythos 5.1. Same weights, different safeguards. Fable 5.1 is generally available on every platform; Mythos 5.1 is gated behind trusted-access programs. If you run coding agents, this release matters more than most. The capability gains concentrate in long-running agentic work, cache reads now cost a quarter of what they did, and a few API changes will break agent harnesses that don't adapt. Here is what changed, what it costs, and how to run Claude Fable 5.1 on your repositories today.

Two names, one model

Start with the naming, because it confuses everyone. Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model. The difference is safeguard configuration: Fable 5.1 ships with production safeguards for general availability, while Mythos 5.1 relaxes specific restrictions for vetted organizations working in cybersecurity and the life sciences.

Fable 5.1 is available now as claude-fable-5-1 on the Claude API, as anthropic.claude-fable-5-1 on Amazon Bedrock, and on Google Cloud and Microsoft Foundry. OpenRouter carries it as anthropic/claude-fable-5.1. Mythos 5.1 is limited to Project Glasswing participants: the Cyber Verification Program for defensive security work and the Life Sciences Verification Program, which Anthropic set up in partnership with the US government. Anthropic's Claude Security product, which scans codebases for vulnerabilities, now runs on Mythos 5.1.

Specs are shared: a 1M token context window at standard pricing across the whole window, 128k max output tokens, and adaptive thinking that is always on. You control reasoning depth with the effort parameter rather than a thinking budget — low, medium, high, and max.

Where the gains landed

The headline numbers, from Anthropic's announcement:

• Terminal-Bench 4.0 (agentic coding): Fable 5.1 scores 55.8%, up from Fable 5's 42.0%. Mythos 5.1, with its reduced cyber-safeguard interventions, reaches 60.9%.
• CursorBench 3.2.0: 73.4% at max effort — the best score measured, ahead of Opus 5 at 70.0%.
• Humanity's Last Exam: 60.9% without tools, 65.0% with tools.
• Terminal-Bench-Science 0.1: 52.6%, against 24.7% for Fable 5.
• OSWorld 2.0 (computer use): 77.9% partial / 41.7% strict.

Artificial Analysis measured 66 on its intelligence index at max effort, the highest score it has recorded, ahead of Claude Opus 5 and Fable 5. Their caveat is worth noting: at max effort it costs about 20% more per task than Fable 5 despite the cache cut, so effort level is a real cost lever, not a nicety.

The qualitative pattern across early-access reports is long-horizon work. Anthropic's page describes a MongoDB engineer who had the model research services code and documentation, design a prototype, then run unattended for hours with verification loops. A Ramp engineer reported a 38-hour run on a machine learning problem that diagnosed a prior result as a label artifact, corrected it, and launched six overnight experiments. These are Anthropic-selected anecdotes, so weight them accordingly — but the benchmark direction agrees with them.

The price cut is in the cache

Token prices are unchanged from Fable 5: $10 per million input tokens and $50 per million output. The cut is elsewhere. Cache reads now cost $0.25 per million tokens — 75% less than before. On other Claude models, a cache read costs a tenth of the base input price; on Fable 5.1 it costs a fortieth.

This matters more for agents than for chat. An agent loop re-reads its full working context — system prompt, tool definitions, conversation history, file contents — on every single turn. That repeated prefix is exactly what the prompt cache serves, and it is where agentic budgets disappear. Anthropic estimates typical workloads get about 25% cheaper relative to Fable 5, and highly agentic, context-heavy work up to 45% cheaper. Batch processing runs $5/$25 per million input/output tokens.

Breaking changes for agent harnesses

If you build on the Messages API — directly or through a harness — four changes need attention. The migration notes are blunt about them.

• Forced tool use is gone. tool_choice of type any or tool returns a 400 error, because a forced call would skip thinking. Keep tool_choice on auto with strict tool schemas, and prompt explicitly when a tool must run.
• Thinking blocks are model-bound in one direction. Fable 5.1 reads earlier models' thinking; no earlier model reads its own. A router that switches models mid-conversation silently drops unreadable blocks.
• History is effectively append-only. Editing the system prompt, tools, or any earlier message invalidates every later thinking block, and the API rejects the replay with "The block is bound to a different conversation". Injected-and-removed per-turn reminders are the classic violation — the fix is turn-scoped system messages (beta) or server-side context editing.
• Behavior shifted in ways that cost tokens quietly: fewer parallel tool calls per turn where batching was only implied, fewer user-facing progress updates during long tool runs, and more whole-file rewrites where a targeted edit would do.

The additive features help agents more than chat: effort can change mid-conversation without invalidating the prompt cache, turn-scoped system messages give per-turn instructions real system-prompt authority, and a display: "updates" mode surfaces progress lines between tool calls while reasoning stays hidden. Output text also carries an invisible watermark, per the EU AI Act — no tokens, no quality impact.

The safety story, briefly

Mythos 5.1 shows the strongest cyber capabilities Anthropic has released, though the company places it within the lower risk category of its Frontier Compliance Framework, and below the next risk tier of its Responsible Scaling Policy for biology. Fable 5.1's own safeguards got more precise rather than just stricter: around 60% fewer interventions per Claude Code session in cybersecurity, explicit permission to use the model for finding software vulnerabilities (exploit development still redirects), and biology safeguards that fire 85% less often on benign medical questions. New API accounts can no longer edit Claude's prior context while preserving its thinking transcripts — a closure of a documented distillation technique.

Running Claude Fable 5.1 in tellnova

tellnova takes any OpenRouter model with a BYO key, and its model picker loads OpenRouter's live catalog — anthropic/claude-fable-5.1 is already in it. Connect an OpenRouter key in Settings, pick the model, and point tellnova at a repository. No update needed.

The fit with background agents is the point:

• Every attempt runs in its own git worktree on its own branch, so a long unattended run never collides with your checkout or another attempt.
• The effort selector maps to the model's effort parameter. Fable 5.1's gains are widest at higher effort, and Artificial Analysis's cost warning is real — pick deliberately instead of pinning max everywhere.
• The 1M context window suits whole-repo reasoning: large refactors, cross-service tracing, dependency upgrades.
• When the work is done, review the diff and ship — tellnova opens the pull request, fork-aware, without touching your main checkout.

Tellnova Cloud runs the same flow on managed E2B sandboxes, with transcripts and usage visible from every signed-in machine. If you are new to subscription-backed models, the walkthrough of using a ChatGPT subscription in tellnova covers the same mechanics for Codex. Grab the app from tellnova.com to try Fable 5.1 against your own repositories.

Should you switch?

If your workload is long agentic sessions with a warm cache — coding agents, research loops, unattended runs — the answer is probably yes: that is the exact profile the cache cut targets, and the capability delta over Fable 5 is largest at high effort. If your usage is short, cache-cold calls, expect costs closer to Fable 5 and decide on capability alone. Either way, re-tune effort consciously, nudge the model to batch independent tool calls, and keep your harness append-only. The full system card is on Anthropic's site if you want the safety detail, and the tellnova GitHub repo if you want to see how the agent side is built.