Claude Fable 5 Guide: The First Mythos-Class Model, Still Served

Claude Fable 5 has the most unusual launch history of any recent frontier model. It shipped, was switched off for every customer three days later, came back three weeks after that, and was succeeded within three months. If you are deciding whether to build on it now, that history matters less than one question: does your integration depend on behaviour that its successor changed?

A launch with an interruption

Anthropic released Claude Fable 5 on 9 June 2026, together with Claude Mythos 5. Anthropic described Fable 5 as a Mythos-class model, a tier above Opus, and the first of that class made safe for general use.

On 12 June the US government applied export controls to both models. Anthropic had no reliable way to verify users' nationality in real time, so it suspended both models for everyone. According to Anthropic, the trigger was a research report describing a way to get around Fable 5's safeguards by prompting it to find software vulnerabilities; Anthropic said its testing showed less capable models could do the same tasks.

The controls were lifted on 30 June and access returned on 1 July, with a new classifier aimed at the reported technique. Anthropic said it blocks that technique in over 99% of cases and acknowledged it also raised false positives on benign coding and debugging requests. Then, on 1 September, Claude Fable 5.1 replaced it as the current Fable model.

Same model as Mythos 5, different safeguards

Fable 5 and Mythos 5 are the same underlying model. Mythos 5 has some safeguards lifted and is restricted to organisations verified through Anthropic's programmes, originally cyber-defence partners in Project Glasswing. Fable 5 is the general-access version.

The practical difference is fallback. Fable 5 runs classifiers covering cybersecurity, biology and chemistry, and distillation. When one fires, the request is handled by Claude Opus 4.8 instead, and the user is told. At launch Anthropic said this happened in fewer than 5% of sessions on average. For most coding and analysis work you will never see it; for security tooling or life-sciences work you will see it often, and the answer you get is an Opus 4.8 answer.

On the API, a classifier decline comes back as HTTP 200 with stop_reason: "refusal". Code that reads content without checking the stop reason first will treat an empty or partial response as a real one.

The specification and the price

Fable 5 has a 1M-token context window and 128K maximum output tokens. Adaptive thinking is always on; the explicit disabled setting and manual thinking budgets both return errors, and depth is controlled with the effort parameter, which defaults to high. Anthropic lists its reliable knowledge cutoff as January 2026.

Pricing is $10 per million input tokens and $50 per million output tokens, with batch at $5 and $25. The catalogue's :batch entry is that discounted asynchronous endpoint for the same model. Cache reads cost $1 per million tokens. Like its successor, it requires 30-day data retention and is unavailable under zero data retention unless Anthropic authorises it.

Fable 5 against its own successor and the Opus line

Anthropic's Fable 5.1 launch table is the cleanest public comparison, and it is not flattering to Fable 5. On Terminal-Bench 4.0 Fable 5 scored 42.0% against 55.8% for Fable 5.1, and against 52.3% for Claude Opus 5. On GDPval-AA v2 it scored 1723, against 1824 for Opus 5. AutomationBench shows 17.1% for Fable 5 and 26.9% for Opus 5.

Two caveats keep this from being a simple ranking. Fable 5 was scored as zero on tasks where its safeguards intervened, which drags its numbers down on agentic benchmarks that touch security. And these are vendor figures; vendor-claimed vs independent benchmarks explains why they tend to run high. Still, the direction is clear: on these measures, Claude Opus 5 at half the per-token price is at least competitive with Fable 5.

Reasons to stay on Fable 5 for now

Fable 5 remains available, and Anthropic has committed not to retire it before 9 June 2027. There are real reasons to keep it for a while.

• Forced tool use. Fable 5 accepts tool_choice of any or a named tool. Fable 5.1 rejects both with a 400.
• Portable thinking blocks. Fable 5.1's thinking blocks can only be read by Fable 5.1 and Mythos 5.1. If your router moves conversations between models, Fable 5 is less disruptive.
• History editing. Fable 5.1 invalidates thinking blocks when earlier turns change, enforced on newer accounts. A harness that rewrites its system prompt or trims messages client-side works on Fable 5 today.
• Pinned evaluations. If you have a validated evaluation set and a regulator or customer who signed off on Fable 5 outputs, a model change is a change-control event regardless of quality.

Against that, the cache price is the strongest reason to move. Fable 5.1 charges $0.25 per million cached tokens against Fable 5's $1. On an agent that re-reads a large cached prefix every turn, that difference can outweigh everything else on the invoice.

Planning the move

A sensible order of work:

1. Grep your code for forced tool_choice and replace it with auto plus strict: true or structured outputs. This works on Fable 5 too, so you can ship it first.
2. Make the conversation append-only: move per-turn instructions into mid-conversation system messages and stop editing earlier messages.
3. Re-run your evaluation on Fable 5.1 and on Opus 5 or Opus 5.5. The comparison you have not run is whether you need the Fable tier at all.
4. Compare cost per completed task, not per request. Prompt caching explained covers why the cache price changes the answer for long sessions.

If the evaluation shows Fable-tier quality matters and your harness is append-only, move to 5.1. If Opus matches it on your tasks, the cheapest migration is down a tier, not across.