Claude Fable 5 in Tarsk

Anthropic shipped Claude Fable 5 on June 9, 2026. It sits above Opus in capability, runs for hours on agentic tasks, and costs twice as much per token. If you delegate multi-file refactors or long research threads in Tarsk, this is the model to try first.

What Anthropic released

Claude Fable 5 is a Mythos-class model made safe for general use. Claude Mythos 5 shares the same weights but drops some safeguards for trusted partners in cybersecurity and, soon, life sciences. Fable is the one you can call from a normal API key.

Both models carry a 1 million token context window, up to 128,000 output tokens, and a knowledge cutoff of January 2026. Anthropic prices them at $10 per million input tokens and $50 per million output tokens, less than half what Mythos Preview cost.

Fable vs Mythos

The names come from the same root: Fable (fabula, "that which is told") and Mythos (Greek for story or account). The model is identical underneath. The difference is which safety classifiers run before a response ships.

Fable 5 adds classifiers for cybersecurity, biology and chemistry, and model distillation. When a request trips one, Anthropic routes the response through Claude Opus 4.8 instead of refusing outright. You get a notice when that happens. Anthropic reports fallback in fewer than 5% of sessions; early testers like Simon Willison note the guardrails fire often enough that the API now exposes explicit fallback signals.

Mythos 5 stays restricted. Cyber defenders in Project Glasswing can upgrade from Mythos Preview today. A broader trusted-access program is coming for security and biomedical research teams that need the biology safeguards removed.

Where it wins

Anthropic positions Fable 5 for work that outlasts a single prompt. Stripe reported a codebase-wide Ruby migration across 50 million lines finished in a day. Cognition's FrontierCode eval puts Fable 5 ahead of other frontier models on hard production coding tasks, even at medium effort.

The model also leads on long-context memory. In Anthropic's Slay the Spire test, file-based notes improved Fable's win rate three times more than they helped Opus 4.8. For vision tasks, Fable 5 completed Pokémon FireRed using only screenshots. Earlier Claude models needed helper harnesses with maps and navigation aids.

Willison spent roughly five hours with Fable on launch day. He had it clone a GitHub repo, swap MicroPython for full CPython in WebAssembly, and ship a working wheel. He then used it to add human-in-the-loop tool approval to Datasette Agent and landed LLM 0.32a3 with new pause-resume APIs. His bill for the day: about $110 on a $100/month Max subscription.

What to expect in Tarsk

Fable 5 is slow and expensive relative to Sonnet or Opus. You pay for depth on hard problems. In Tarsk, that usually means agent threads with many tool calls: reading files, running checks, iterating on a plan across dozens of steps.

• Reasoning effort from low through max, so you can cap spend on simpler subtasks.
• 1M-token context for whole-repo analysis without aggressive trimming.
• Tool calling tuned for long-horizon agent workflows.
• Image and PDF inputs when you need screenshots or documents in the same thread.

Reserve Fable for migrations, multi-service debugging, and research threads where Opus stalls. Use Sonnet or a smaller model for quick edits so token spend stays predictable.

Providers in Tarsk

Eleven Tarsk providers list Claude Fable 5 today. Mythos 5 does not appear in public catalogs yet because Anthropic limits it to approved partners.

First-party and cloud

• Anthropic: claude-fable-5 at console.anthropic.com.
• Amazon Bedrock: us.anthropic.claude-fable-5, eu.anthropic.claude-fable-5, global.anthropic.claude-fable-5.
• Azure: claude-fable-5 through Azure AI Foundry.

Aggregators and gateways

• OpenRouter: anthropic/claude-fable-5 and ~anthropic/claude-fable-latest (auto-tracks the newest Fable release).
• ZenMux: anthropic/claude-fable-5.
• Vercel AI Gateway: anthropic/claude-fable-5.
• Cloudflare AI Gateway: anthropic/claude-fable-5.
• OpenCode Zen: claude-fable-5.

Other routes

• Venice AI: claude-fable-5 (slightly higher list pricing).
• FreeModel: claude-fable-5.
• GitLab Duo: duo-chat-fable-5 for Agentic Chat on GitLab Premium and Ultimate.

Try it in Tarsk

1. Add a provider key under Settings → Providers. Anthropic or OpenRouter are the fastest paths.
2. Open Settings → Models, search fable, and enable the endpoint you want.
3. Select that model in your thread and send a task with tools enabled.

Start at medium reasoning effort. Raise it when the agent needs to plan across many files or recover from a failed test run. Watch token usage in the thread header so a long session does not surprise you at billing time.

Summary