Model event

Claude Fable 5 for agent loops

Anthropic's first Claude 5 model launched, disappeared under export controls, and came back — with refusal handling, fallbacks, effort, and task budgets as API primitives. Here is the verified timeline and what it changes for loop design.

Answer-first summary

Claude Fable 5 is Anthropic's most capable generally available model, built for long-horizon agentic work: it launched June 9, 2026, was suspended June 12 under US export controls, and has been back globally since July 1 on the Claude API, Claude.ai, Claude Code, and Claude Cowork.

For loop builders, the headline is not the benchmarks. It is that four things every production loop already needed — refusal handling, model fallback, per-turn spend control, and task budgets — are now explicit API primitives on Fable 5 rather than things you improvise.

Verified timeline

Date	What happened
2026-06-09	Claude Fable 5 and Claude Mythos 5 announced. Fable 5 generally available on the Claude API (`claude-fable-5`), AWS, Google Cloud, and Microsoft Foundry.
2026-06-12	US government applies export controls to both models; Anthropic restricts access to foreign nationals. Trigger: an Amazon research report on a method to bypass Fable 5's safeguards via software-vulnerability prompts.
2026-06-30	Export controls on Fable 5 and Mythos 5 lifted.
2026-07-01	Fable 5 returns globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. Pro/Max/Team plans include it up to 50% of weekly usage limits through July 7, then usage credits.

Key facts

Fact	Value
API model ID	`claude-fable-5` (Mythos 5: `claude-mythos-5`, Project Glasswing only)
Context / output	1M-token context window by default; up to 128k output tokens per request
Pricing	$10 / M input tokens, $50 / M output tokens
Thinking	Adaptive thinking always on; depth controlled by the `effort` parameter; raw chain of thought never returned
Refusals	Safety classifiers can decline requests; the API returns `stop_reason: "refusal"` as a successful HTTP 200, with the classifier identified
Fallback	Server-side `fallbacks` parameter (beta), SDK middleware, or manual retry on another Claude model; fallback credit refunds the prompt-cache cost of switching
Task budgets	Supported at launch (beta header `task-budgets-2026-03-13`), alongside the memory tool, compaction, and context editing
Data retention	30-day retention required; not available under zero-data-retention

What Fable 5 changes for loop design

Loop field	What is new on Fable 5
Stop condition	A refusal is now a first-class stop signal: `stop_reason: "refusal"` arrives as HTTP 200, not an error. A loop that only catches exceptions will read a refusal as a successful turn and keep going. Treat it as a blocker stop.
Budget	Two native levers: the `effort` parameter caps thinking spend per turn, and task budgets (beta) cap spend across a whole task. Loop budgets no longer have to be homemade token counters.
Degraded mode	The `fallbacks` parameter retries a refused request on another Claude model server-side. Decide per loop whether a fallback model may complete the leg or whether the loop should stop for review — a security-scanning loop and a docs loop want different answers.
State	A 1M-token window, the memory tool, and compaction stretch how long one leg can run — but durable state in files, PRs, and logs is still what survives the loop, not the context window. Longer legs raise the value of checkpoints, not lower it.
Approval boundary	Anthropic says Fable 5 "can work autonomously for longer than any previous Claude models." Longer unattended runs mean the approval gates and verifiers carry more weight, not less. The leash does not get shorter because the dog got smarter.

What is not established

How long "longer than any previous Claude model" is — there is no published duration benchmark; treat vendor autonomy claims as direction, not spec.
The technical details of the safeguard bypass that triggered the export directive — Anthropic has not published them.
Post-introductory usage-credit economics on subscription plans, which vary by plan and are still settling.

Sources

Timeline and API facts verified against Anthropic's announcements and platform documentation on 2026-07-13.

FAQ

Is Claude Fable 5 available right now?

Yes. Fable 5 launched June 9, 2026, was suspended June 12 under US export controls, and returned July 1, 2026 on the Claude API, Claude.ai, Claude Code, and Claude Cowork after the controls were lifted June 30.

Why was Fable 5 taken down?

The US government applied export controls to Fable 5 and Mythos 5 on June 12, 2026, requiring Anthropic to restrict access to foreign nationals. Anthropic says the trigger was an Amazon research report describing a method to bypass Fable 5's safeguards by prompting it to identify software vulnerabilities. The controls were lifted June 30.

What is the difference between Claude Fable 5 and Claude Mythos 5?

Same underlying model. Fable 5 is the generally available version with safety classifiers that can decline certain requests; Mythos 5 ships without those classifiers and is limited to approved organizations through Project Glasswing.

Does Fable 5 change how agent loops should be designed?

The anatomy is unchanged — trigger, objective, verifier, budget, stop condition, approval gate. What changes is that several loop fields now have first-class API primitives: refusals arrive as a stop_reason, fallbacks and task budgets are parameters, and effort levels control per-turn spend.

What does Claude Fable 5 cost?

$10 per million input tokens and $50 per million output tokens on the API, with a 1M-token context window and up to 128k output tokens per request. On Claude subscription plans it runs on usage credits after an included introductory window.