Claude Fable 5

Anthropic's highest-capability model in the Claude 5 generation. The Fable name marks the first departure from the Haiku / Sonnet / Opus tier structure that defined the Claude 3 and 4 families. claude-sonnet-5 handles the agentic default role; Fable 5 is the ceiling — positioned for the most complex reasoning, analysis, and long-horizon tasks where quality outranks cost.

Positioning

The structural move from Opus to Fable is deliberate. Where Opus implied scale (larger, heavier, more expensive), Fable implies something different: coherent, structured reasoning over long contexts with consistent internal logic. The name signals a shift in how Anthropic frames the top-tier model — not as a compute-maxed version of Sonnet, but as a model built for extended analytical work that requires holding complex ideas in relation across a long session.

In practice: use Fable 5 where claude-opus-4-8 was previously the default — complex legal/financial agent work, elaborate multi-step reasoning tasks, and sessions where failure to maintain coherence across a long context has meaningful consequences.

Context Window

1M tokens. Max output: 128K tokens.

Thinking

Supports extended-thinking in adaptive mode. The model determines reasoning depth based on task complexity. Extended thinking is active by default at high effort.

Strengths

• Complex reasoning: Sustained coherence across long, multi-step analytical tasks where mid-tier models lose the thread.
• Long-horizon agent work: Maintains task context and goals across extended autonomous sessions without drift.
• Analytical depth: Best-in-class for tasks requiring integration of many information sources with consistent logical structure.

Weaknesses

• Higher cost than claude-sonnet-5 — overkill for routine coding, summarization, or high-volume pipelines.
• No manual budget_tokens control — adaptive thinking only, consistent with Opus 4.8.

Use Cases

Best for: complex financial or legal agent benchmarks, extended reasoning tasks, high-stakes analysis where coherence failures are costly, research synthesis across many sources.

Not ideal for: routine agentic coding (use claude-sonnet-5), tasks where cost-per-token matters at volume.

Availability

Claude API, AWS Bedrock, Google Vertex AI, claude.ai (Max plans), claude-code.

Related

claude-sonnet-5 · claude-opus-4-8 · extended-thinking · agentic-workflows · claude-code

Sources

• Claude models overview — Anthropic docs
• Anthropic pricing