Claude Fable 5

The First Claude 5 Model Above the Standard Opus Tier

Anthropic launched Claude Fable 5 on June 9, 2026 as its most capable generally available model at the time. It also introduced the broader Mythos-class tier: Claude Mythos 5 shares the same underlying capabilities and specifications but is available only to approved Project Glasswing customers and does not use Fable’s additional safety classifiers.

Fable 5 was designed for problems that were previously too complex, long-running, or ambiguous for earlier Claude models. Anthropic’s prompting guidance specifically highlights work that may take a person hours, days, or weeks to complete.

The model combines a 1,000,000-token context window, up to 128,000 output tokens, text and image input, always-on adaptive thinking, and tool-oriented agent workflows. Anthropic classifies comparative latency as slower, which matches its role as a premium model for difficult tasks rather than routine low-latency calls.

Today, Fable 5 remains available as an active legacy model. Anthropic recommends considering Fable 5.1 for improved performance, so Fable 5 is especially useful as a production baseline for teams already running it or evaluating what changed between the original and updated Fable generation.

• Released June 9, 2026.
• Built for demanding reasoning and long-horizon agentic work.
• 1M context window and 128K maximum output.
• Text and image input with text output.
• Active legacy model with migration path to Fable 5.1.

Production Integrations Need a Refusal Path

Anthropic’s documentation calls out refusal handling as a headline integration change for Claude Fable 5. A declined request can return stop_reason: "refusal", so applications should not assume every successful HTTP request ends with a normal completion.

The classifiers target areas including offensive cybersecurity, certain biology and life-sciences content, and attempts to extract summarized thinking. Anthropic also notes that benign cybersecurity or beneficial life-sciences requests can trigger these safeguards in some cases.

This means production systems should explicitly handle refusal responses. Depending on the workload, the application may show a refusal to the user, route the request for human review, or retry on another approved Claude model through a configured fallback path.

Billing rules also matter. Mid-stream refusals can bill the input and output already generated. Anthropic introduced fallback-credit behavior to avoid charging prompt-cache cost twice when a declined Fable request is retried on another model.

Prompt Caching Is Important for Long Agent Sessions

A long-running agent may repeatedly reuse system instructions, repository context, policy documents, research material, or other stable prefixes. Without caching, those tokens are billed as standard input each time they are processed.

Anthropic prices 5-minute cache writes at $12.50 per million tokens, 1-hour cache writes at $20 per million tokens, and cache reads at $1 per million tokens for Fable 5.

That cache-read rate is one of the clearest economic differences between Fable 5 and Fable 5.1. The newer model keeps the same $10/$50 base pricing but lowers cache reads to $0.25 per million tokens.

Batch API workloads receive a 50% discount on standard input and output. For offline evaluations, large research jobs, or bulk processing that does not require interactive latency, batching can materially change the economics.

• Standard input: $10.00 per 1M tokens.
• Output: $50.00 per 1M tokens.
• 5-minute cache write: $12.50 per 1M tokens.
• 1-hour cache write: $20.00 per 1M tokens.
• Cache read: $1.00 per 1M tokens.
• Batch API: 50% discount on standard input and output.

Fable 5 Changed the Thinking API Contract

Unlike some earlier Claude models, Fable 5 does not allow thinking to be disabled. Passing thinking: {"type": "disabled"} returns a 400 error. Manual extended-thinking budgets with thinking: {"type": "enabled", "budget_tokens": N} are also not supported.

The default effort level is high. Lower effort can be useful for less demanding steps where evaluation shows that full compute is unnecessary.

Raw chain-of-thought content is never returned. Anthropic’s API can provide summarized thinking or omit readable thinking content depending on the thinking.display setting. In multi-turn tool loops, thinking blocks should be passed back unchanged according to Anthropic’s API requirements.

This means migration from older Claude models requires both quality testing and API compatibility testing. A prompt that worked before may still be valid, while the surrounding thinking configuration may need to change.

Built for Stateful Workflows

Anthropic lists support for effort controls, task budgets in beta, the memory tool, code execution, programmatic tool calling, context editing for tool-result clearing, compaction, and vision.

These features are especially relevant to agents because long tasks need active state management. Tool results can accumulate, old context can become less useful, and the application may need to preserve durable memory while compacting transient details.

Fable 5 also uses a tokenizer introduced with Claude Opus 4.7. Anthropic notes that the same text can produce roughly 30% more tokens than on models released before Opus 4.7, depending on content and workload shape. Cost comparisons with older Claude generations should therefore use measured token counts rather than assume prompt lengths translate directly.