Claude Fable 5: What It Is and When to Use It

Claude Fable 5 launched on June 9, 2026, as Anthropic's most capable publicly available model. It is the first model from Anthropic's Mythos class, a capability tier that sits above the Opus family, built for long-horizon tasks that previous models could not sustain.

A note on timing: within three days of launch, a US government export-control directive forced Anthropic to suspend access globally. That suspension was lifted on July 1, 2026. Fable 5 is fully available again across Claude.ai, the Claude API, AWS, Google Cloud, and Microsoft Foundry.

This post cuts through the launch noise to answer the questions that actually matter for developers: what tasks is Fable 5 genuinely better at, what does it cost, how does the safety fallback affect your integration, and when should you stick with Opus 4.8 instead.

Fable 5 is not the next Opus update. The architecture is fundamentally different in how it approaches long tasks. Four changes matter most for developers.

All numbers below come from Anthropic's official launch announcement on June 9, 2026, comparing Claude Fable 5 to Claude Opus 4.8.

Starred benchmarks represent Mythos 5 results, a sibling model with the same capabilities. On cybersecurity-related benchmarks, Fable 5 scores closer to Opus 4.8 because its safety classifiers route those queries elsewhere. For non-cybersecurity tasks, Fable 5 and Mythos 5 perform within 1 to 3 percentage points of each other.

Third-party evaluations at launch pointed the same direction. Hex reported Fable 5 was the first model to score 90% on its core analytics benchmark for complex, long-running analytical tasks. Genspark said Fable 5 beat every other model in its evaluations, with particularly strong UI design and game coding results. Wharton professor Ethan Mollick, who had early access, found it capable of working autonomously for up to 12 hours on multi-page specifications.

This is Fable 5's clearest advantage. Large-scale migrations, repository-wide refactors, complex implementations that need repeated test-and-fix loops, and multi-day autonomous sessions are exactly what the model was designed for.

The most concrete example from launch week: Stripe reported that Fable 5 compressed a codebase-wide migration across a 50-million-line Ruby infrastructure into a single day. The team estimated the same work would take a group of engineers more than two months by hand. The task required semantic understanding of thousands of files, correct application of transformations in unusual edge cases, and autonomous verification that each change did not break anything downstream.

If you run these jobs through an agent harness, the practical setup matters as much as the model. Our Claude Code cheatsheet covers running long-horizon coding sessions, and the multi-agent AI coding workflow guide covers delegating across parallel sub-agents, which is where Fable 5 pulls furthest ahead.

For shorter coding tasks, such as writing a unit test, explaining a snippet, or generating a simple component, the gap between Fable 5 and Opus 4.8 shrinks considerably. The premium is harder to justify there.

Research synthesis across dozens of sources, multi-stage analysis feeding into a final deliverable, and projects that teams hand off and review as completed work rather than supervising step by step. Fable 5's larger context window and stronger sustained reasoning make it more reliable across this kind of work than Opus 4.8.

Fable 5 uses vision to understand diagrams, charts, and tables nested in PDFs and documents, not just plain-text extraction. For contract review, financial report analysis, and architecture documentation, it can reason over the full document structure rather than working from text alone. Multiple law firms reported at launch that in blind review, Fable 5's redlines matched or beat their existing model on every comparison.

Agent harnesses where the model must plan across stages, use tools, check intermediate results, change direction when an approach is not working, and complete the full job without human supervision in between. Fable 5 is more reliable at maintaining coherent state and goals across long chains of tool calls than earlier models. When you run several of these agents at once, git worktrees for parallel AI coding agents keep each session isolated on its own branch.

Fable 5 can write its own tests, run them, check the results, and use vision to compare its output against a design spec or goal image. This changes the economics of code review for high-value tasks: the model provides a validated result rather than one that still needs full human verification.

This is the section most launch-week reviews handle poorly. The fallback behavior is genuinely novel and has real integration implications.

What triggers it. Fable 5 includes safety classifiers that watch for certain query types, primarily in three categories: offensive cybersecurity (exploit development, agentic hacking), biology with dual-use potential, and chemistry. When a query trips a classifier, Fable 5 does not return an error. It reroutes the request to Opus 4.8 and gives you the Opus 4.8 response instead.

What the API response looks like. When a fallback occurs, the API returns a standard HTTP 200 response with stop_reason: "refusal" in the message metadata, plus a stop_details object naming the classifier category that triggered it. This is different from a traditional refusal that returns no useful content: you still get a usable answer, just from Opus 4.8 rather than Fable 5.

Three practical changes for your integration. First, treat stop_reason: "refusal" as a success case, not an error. Your existing handling probably does not cover it if you have only used earlier Claude models. Second, decide whether fallback responses need to be flagged differently in your UX; for most enterprise applications a high-quality Opus 4.8 answer is still a good outcome. Third, on the raw API the fallback is opt-in through a fallbacks parameter, so make sure your request enables it rather than assuming a refusal is automatically rerouted.

How often does it trigger? Anthropic reports the classifiers fire in fewer than 5% of sessions on average. For data analysis, code review, content generation, and general knowledge work, the restriction will not appear at all. The covered categories are narrow and targeted.

Billing. You are not charged Fable 5 prices for rerouted requests. A fallback response is billed at Opus 4.8 rates, and a request declined before any output is not billed at all.

Fable 5 costs twice as much as Opus 4.8 at every tier. The right way to think about whether that premium pays off is cost per completed task, not cost per million tokens.

If Fable 5 completes a complex migration in one attempt and Opus 4.8 fails twice before succeeding on the third, you have spent triple the Opus cost before getting a result. Even at double the per-token rate, Fable 5 can be the cheaper option when first-attempt success rates differ significantly.

Concretely: suppose a difficult task uses 500,000 input tokens and 100,000 output tokens. That is one Fable 5 run at about $10. If a cheaper model costs $4 per attempt and needs three attempts to produce a usable result, Fable 5 comes out ahead. Input tokens served from the prompt cache receive up to a 90% discount on the input price, which brings the effective cost down further for applications with large system prompts or repeated context.

Mandatory data retention. Fable 5 requires a 30-day data retention window on all API traffic, even for organizations that previously had zero-retention agreements. A request from an organization configured for zero retention returns a 400 error on every call. Anthropic states retained data is not used for training and is deleted after 30 days in most cases, with exceptions for safety investigations and legal obligations. For regulated industries, this needs to clear your compliance review before you migrate.

Model ID and availability. Use claude-fable-5 as the model string in API calls. Fable 5 is available on the Claude API, Claude.ai (Pro, Max, Team, and Enterprise), Amazon Bedrock, Google Cloud, and Microsoft Foundry.

US-only inference. If your workload must run on US infrastructure for compliance reasons, US-only inference is available at a 1.1x multiplier on standard pricing.

The practical approach most teams are using: route high-complexity, high-stakes, long-horizon jobs to Fable 5, and default routine traffic to Opus 4.8 or a smaller model. This hybrid gets frontier-quality results where they matter without paying frontier prices everywhere.