Claude Fable 5

Overview. Claude Fable 5 is Anthropic's proprietary, closed-weight Mythos-class reasoning model, generally available since June 2026. It runs with mandatory, always-on adaptive reasoning and a one-million-token context window, accepting text and image input and returning text output. Developers reach for it when a task needs sustained, multi-step reasoning over large amounts of context rather than a fast conversational reply.

Key strengths. Artificial Analysis benchmarking places Fable 5 in frontier territory, with a Coding Index of 76.5 and an Intelligence Index that has moved up to 62.1 in the most recent reading, alongside an Agentic Index of 56.6. The model exposes native tool calling, structured output, and JSON mode, which makes it well suited to agentic pipelines that call functions and expect schema-constrained responses. Its five-tier reasoning effort control (max, xhigh, high, medium, low) lets a team trade latency for depth on harder problems.

When to use it. Use Fable 5 for long-context code review, multi-document research synthesis, and agentic workflows that chain tool calls and need structured JSON back. The one-million-token window and 128,000-token output ceiling suit large codebases or research corpora better than short back-and-forth chat. Because reasoning is always on and tuned for depth, it fits batch and asynchronous jobs more naturally than latency-sensitive, real-time interactions.

Limitations. Reasoning cannot be turned off and the model does not accept a temperature parameter, so it isn't a drop-in swap for lightweight conversational deployments that expect fast, low-effort responses. Anthropic has not disclosed parameter count, architecture details, or a training data cutoff for this model, and input support stops at text and image — there is no audio input or output. Its reasoning-first design points toward a batch-style latency profile rather than real-time chat.

Sovereignty note. Anthropic is a US-headquartered creator, so Fable 5 Inference carries CLOUD Act exposure rather than EU-residency protections; our research did not surface a published training-on-input opt-out or EU-specific data residency commitment for this model. It is a closed-weight, proprietary model governed by Anthropic's Acceptable Use Policy.