Claude Fable 5 Is Back: What's New in Anthropic's Most Capable Model

Anthropic's most capable widely released model is back online. Claude Fable 5, along with its limited-release sibling Claude Mythos 5, returned to users on July 1, 2026, following a temporary suspension tied to a U.S. government export control review. Access is now fully restored across the Claude API, Claude.ai, Claude Code, Claude Cowork, and major cloud platforms.

Claude Fable 5 was originally released on June 9, 2026, as Anthropic's fifth-generation model, built for the most demanding reasoning tasks and long-horizon agentic work, the kind of multi-day, multi-stage projects earlier models couldn't sustain. Three days later, the model was pulled after a security researcher found a way to prompt it into surfacing exploitable software vulnerabilities. Anthropic worked with the government to resolve the issue, and access was fully reinstated once the export controls were lifted.

The headline change for anyone building on Fable 5 is new safety classifiers that can decline requests outright. Unlike a typical error, a declined request returns a standard successful response with a specific "refusal" status, along with which classifier triggered it. Claude Mythos 5 does not include these classifiers and remains available only to approved organizations through Anthropic's Project Glasswing.

For developers, that means three things to plan for: how your integration handles a refusal response, how it falls back to another Claude model when a request is declined, and how billing works when that happens. Anthropic offers both a server-side option that retries automatically and a client-side SDK middleware for TypeScript, Python, Go, Java, and C#. Notably, you're not billed for a request that gets refused before any output is generated, and switching models on a retry doesn't cost you the prompt-cache fee twice.

On raw capability, nothing has been scaled back. Claude Fable 5 ships with a 1-million-token context window by default and supports up to 128,000 output tokens per request. Pricing holds at $10 per million input tokens and $50 per million output tokens, with the usual prompt-caching discount still in place. The model also supports adaptive thinking, task budgets, the memory tool, code execution, programmatic tool calling, context editing, compaction, and vision, the full toolkit for building serious agentic applications.

Availability-wise, Claude Fable 5 is generally available now on the Claude API, the Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry. For everyday use, it's live across Claude.ai, Claude Code, and Claude Cowork, and is included for Pro, Max, Team, and select Enterprise plans for up to 50% of weekly usage limits through July 7, after which it shifts to usage credits.

If you were mid-migration from Claude Opus 4.8 or an earlier Mythos preview when the suspension hit, Anthropic's migration guide walks through the upgrade path step by step, including how the new refusal and fallback behavior changes what your code needs to handle.

The takeaway: Claude Fable 5 returns with the same frontier-level reasoning it launched with, plus a more robust safety layer that's now a first-class part of the API rather than an afterthought. If your team paused a rollout during the suspension, this is a good moment to pick it back up.