The questions everyone's asking

Availability

Is Fable 5 back?

Yes. Fable 5 was suspended worldwide on June 12, 2026 under a US government export-control directive. The Commerce Department lifted the controls on June 30, and Anthropic restored Fable 5 globally on July 1 — Claude API, claude.ai, Claude Code, and Claude Cowork, with AWS, Google Cloud, and Microsoft Foundry re-enabling in progress. It returned with a new safety classifier. The full story: Fable 5 returns.

Is Claude Fable 5 down?

No — not anymore. It came back July 1, 2026 and is live on the API, claude.ai, Claude Code, and Claude Cowork. If it's missing on AWS, Google Cloud, or Microsoft Foundry, that's re-enablement lag, not another suspension. Check every surface in real time on our status page.

How do I get access?

Three doors: the API (claude-fable-5), Claude Code (it's the engine), or claude.ai paid plans during the window. AWS, Google Cloud, and Microsoft Foundry are re-enabling access following the July 1 restoration. Walkthroughs in Start.

What's the exact model ID?

claude-fable-5 — no date suffix exists. Bedrock prefixes it: anthropic.claude-fable-5.

Did the model ID change after the return?

No. Still claude-fable-5, exactly as at launch. Any config or script that kept the ID through the suspension started working again on July 1 with zero changes — nothing to migrate, nothing to rename.

The shutdown

Why was Fable 5 banned?

A US government export-control directive, June 12, 2026 — issued after Amazon researchers reported a jailbreak that got Fable 5 to flag software vulnerabilities and write exploit-demo code. The order required restricting foreign nationals; with no way to verify nationality in real time, Anthropic pulled access entirely, for everyone. Commerce lifted the controls June 30, citing Anthropic's coordination commitments. Day-by-day chronology: the timeline; Anthropic's own account: what Anthropic said.

Was Fable 5 actually dangerous?

Anthropic disputed the finding as narrow, not universal. Its June 16 testing found other models — including Opus 4.8, GPT-5.5, and Kimi K2.7 — could identify the same vulnerabilities, and every model tested could reproduce the report's one demonstration. Commerce's CAISI later tested the prior and new safeguards and called them “extraordinarily strong.” Anthropic's standing caveat: no model can be made fully robust to jailbreaks. Background: how the controls got lifted.

Costs

Is Claude Fable 5 free?

On claude.ai paid plans (Pro, Max, Team, select Enterprise): included at 50% of weekly usage limits through July 7, 2026. After July 7 it takes usage credits. The API is never free — $10/$50 per million tokens. There is no free public API tier.

How do I use it without paying extra?

Ride the window: through July 7, any Pro, Max, Team, or select Enterprise plan includes Fable 5 at half of weekly limits with nothing extra to buy. Pick it in the claude.ai model picker, or /model fable in Claude Code. Spend it on the hard tasks — what to run on it — and let cheaper models do the rest.

How long is the return window?

July 1 – July 7, 2026, at half of weekly usage limits on Pro, Max, Team, and select Enterprise plans. The original launch offer was free through June 22 — fourteen days, of which subscribers got about three before the suspension — so the make-good is shorter and shallower than what was promised, and some subscribers have said so loudly.

What happens after July 7?

Subscription use moves behind usage credits on July 8. Standard Enterprise is credits-only throughout. API access never changes — same $10/$50, same model ID. How credits work, in detail: the credits page.

How much does it cost on the API?

$10 per million input tokens, $50 per million output — twice Opus 4.8. It's also more token-efficient (Notation Capital measured a third of the reasoning tokens on physics work), so judge by completed-task cost. Caching cuts repeat input to ~10%.

Behavior

Why does Fable refuse my coding request — or hand it to Opus 4.8?

The new classifier's "expanded safety margin" deliberately fires on some benign requests, and Anthropic admits it produces false positives during routine coding and debugging. In some flows the model hands the task to Opus 4.8 and notifies you. It's a trade-off Anthropic made to get the model back; details in our classifier piece.

Does it refuse things?

Three classifiers (cyber, bio/chem, distillation) fall back to Opus 4.8 when tripped. Since the July 1 return there is also a new safety classifier aimed at the reported cyber jailbreak — it blocks the technique in over 99% of cases, at the cost of more false positives on routine coding. Separately — and controversially — frontier-AI-development tasks can be quietly limited without notice.

Why is Fable 5 slow?

By design. At higher effort it takes longer turns, spending more reasoning per step so hard tasks land right the first time — one-shot depth over fast back-and-forth. Dial it down with output_config.effort for snappier replies, or route quick tasks to a cheaper model. The effort dial, explained: Effort; prompting for it: Prompting Fable 5.

Why is my old code 400ing?

Sampling params, thinking budgets, and prefills are gone; explicit thinking: disabled is Fable-specific 400 — omit the field. The full list with fixes: gotchas.json / Recipe 01.

Context window? Output limit?

1M tokens in, up to 128K out (stream anything large). No long-context premium.

The lineup

What's the difference between Fable 5 and Mythos 5?

The same underlying model, two doors. Fable 5 is the general release with strong safeguards — "a Mythos-class model made safe for general use." Mythos 5 has safeguards lifted in narrow domains and is restricted to vetted partners: restored for US organizations on June 26, 2026, with broader domestic and international expansion running through Project Glasswing. Full story: Mythos, explained.

Is it better than Opus 4.8?

On the hardest work, clearly — 91/100 on Every's senior-engineer exam vs Opus's 63. For daily work, Opus at half price is often the right call. The routing table: Fable vs Opus 4.8.

Is Fable 5 better than Sonnet 5?

Different jobs. Fable 5 is the frontier model ($10/$50) for the hardest long-horizon work; Sonnet 5, launched June 30, is the budget agent runner — $2/$10 intro through Aug 31, then $3/$15 — and the new default on Free and Pro. On agentic coding Sonnet 5 scores 63.2% vs Opus 4.8's 69.2%: great for volume agent fleets, not a substitute on frontier problems. Head-to-head: Fable vs Sonnet 5.

What is Fable 5 best at?

Four things, reliably: long-horizon autonomy (multi-hour agent runs that stay coherent), first-shot complex builds, whole-codebase review, and vision. If Opus 4.8 or Sonnet 5 can do the task, use them and save the window. The task-by-task routing guide: What to run on Fable 5.

Moral: most questions about a new model are really one question — "is it worth my time yet?" For hard work: yes, and through July 7 the return window makes finding out cheap.