Claude Fable 5: Anthropic's Most Capable Model Yet

Anthropic just raised its own ceiling. Claude Fable 5 is the company’s most capable model to date — and unlike a routine point release, it’s aimed squarely at the hardest, longest-running work: autonomous agents that plan, build, and verify across many steps without a human nudging them every few minutes.

If you’ve been following the state of AI coding assistants, this is the model built for the frontier that conversation keeps pointing at — not snappier autocomplete, but agents you can hand a whole task and trust to run.

A new tier above Opus

Anthropic’s lineup has long followed three names: Haiku for speed, Sonnet for balance, and Opus for raw capability. Fable 5 introduces a new name that sits above the Opus tier as the flagship for frontier reasoning and agentic execution. Opus 4.8 remains an excellent top-of-Opus model, and Sonnet 4.6 is still the workhorse for everyday tasks — but when a job genuinely needs the most capable model available, Fable 5 is now that model. (If the naming is confusing, our guide to Anthropic’s 2026 model lineup maps the whole family.)

What’s actually new

• Always-on reasoning. Fable 5 thinks before it answers by default — there’s no manual “thinking budget” to tune. Instead you set an effort level (from low to max) to dial how hard it works against the cost you’re willing to pay. The raw chain of thought isn’t exposed; you get a readable summary if you ask for one.
• Built for long-horizon autonomy. This is the headline. A single request can run for many minutes as the model gathers context, makes changes, and checks its own work. It’s notably stronger at sustained multi-step tasks, coordinating parallel sub-agents, and keeping notes in a memory file it can consult later.
• A 1M-token context window with up to 128K tokens of output — room to hold a large codebase or document set and still produce a substantial result.
• The same tokenizer as Opus 4.8, so token counts carry over cleanly if you’re already on that model.

What it costs

Frontier capability comes at a frontier price. Fable 5 sits above the Opus tier on cost, so it’s a deliberate choice, not a default.

Model	Input ($/1M tokens)	Output ($/1M tokens)
Claude Fable 5	$10.00	$50.00
Claude Opus 4.8	$5.00	$25.00
Claude Sonnet 4.6	$3.00	$15.00
Claude Haiku 4.5	$1.00	$5.00

The practical read: reach for Fable 5 when a task’s value clearly justifies the spend — a complex migration, deep research, an overnight autonomous build. For most day-to-day work, Sonnet 4.6 or Opus 4.8 deliver excellent results for a fraction of the cost.

Safety and deployment notes

A few things teams should know before wiring Fable 5 into production:

• Safety classifiers can decline a request. Fable 5 isn’t intended for sensitive biology or most cybersecurity work, and a classifier may return a “refusal” instead of an answer. Benign adjacent tasks occasionally trip it, so production apps should handle that case — typically by falling back to another model such as Opus 4.8.
• It requires 30-day data retention. Fable 5 isn’t available to organizations configured for zero data retention. Check your account’s retention settings before you build against it.
• Mythos 5 is the same model, different door. Anthropic also offers Claude Mythos 5 — identical capabilities and pricing — through its Project Glasswing program. Unless you’re in that program, claude-fable-5 is the model you want.

When to use it (and when not to)

Use Fable 5 for complex agentic builds, deep multi-source research, long autonomous runs, and the hardest reasoning problems — the work where one more increment of capability changes the outcome.

Skip it for simple chat, classification, summarization, or high-volume pipelines. Those don’t need a frontier model, and the cost difference is real. Sonnet 4.6 is the balanced default; Haiku 4.5 is the budget speedster.

The takeaway

Fable 5 is a bet that the next leap in AI usefulness isn’t a better single answer — it’s a model you can trust to work on its own for a long stretch and come back with something done. That only pays off if you meet it halfway: a clear spec up front, real verification on the back end, and guardrails around anything destructive. The meta-harnesses and agent platforms maturing alongside it exist for exactly that reason. Give the model a goal and a way to check its work, and Fable 5 is the most capable partner Anthropic has shipped.