Claude Fable 5 vs Claude Sonnet 5: pricing, Quality, Value, and benchmarks

Buyer access

Pricing, app access, and Conversation Value

• Humanity's Last Exam

  Normalized quality input

  53.3%

  Artificial Analysis Claude Fable 5 evaluation | Independent Fable 5 production-config result with adaptive reasoning, maximum effort, and safeguard fallback enabled.

• AA-Omniscience

  AA-Omniscience Index

  40.1

  Artificial Analysis AA-Omniscience evaluation | Independent Artificial Analysis result for Claude Fable 5 (Adaptive Reasoning, Max Effort, safeguard fallback enabled). Display-only; this row does not affect Quality or Value scores.

• SWE-Bench Pro

  Normalized quality input

  80.0%

  Anthropic Claude Fable 5 and Mythos 5 system card | Anthropic vendor-reported Fable 5 result averaged over five trials. Use the Fable 5 SWE-Bench Pro column, not Mythos 5 or SWE-Bench Verified.

• Terminal-Bench 2.1

  Agentic terminal task completion

  84.3%

  Anthropic Claude Fable 5 and Mythos 5 system card | Anthropic vendor-reported Fable 5 Terminal-Bench 2.1 result using mini-SWE-agent at high effort; 20.9% of trials used safeguard fallback routing.

• Humanity's Last Exam

  Normalized quality input

  43.2%

  Anthropic Claude Sonnet 5 system card | Vendor-reported Anthropic Sonnet 5 HLE no-tools score. Do not replace with tools-enabled HLE variants.

• SWE-Bench Pro

  Software engineering task resolution

  63.2%

  Anthropic Claude Sonnet 5 system card | Vendor-reported Anthropic Sonnet 5 SWE-Bench Pro score. Do not substitute SWE-Bench Verified.

• Terminal-Bench 2.1

  Agentic terminal task completion

  80.4%

  Anthropic Claude Sonnet 5 system card | Vendor-reported Anthropic Sonnet 5 Terminal-Bench 2.1 score using mini-SWE-agent. Display-only companion evidence.

• AA-Omniscience

  AA-Omniscience Index

  15.3

  Artificial Analysis AA-Omniscience evaluation | Independent Artificial Analysis result for Claude Sonnet 5 (max). Display-only; this row does not affect Quality or Value scores.

The most intelligent model in the current field, but its record capability comes with roster-leading token prices, quarter-rate token throughput, and safeguard-triggered fallback routing.

Claude Fable 5 is the most intelligent generally available model in the current field: its production configuration leads the independent Artificial Analysis Intelligence Index, while Anthropic reports 80.0% on SWE-Bench Pro and 84.3% on Terminal-Bench 2.1. Buy it selectively, not by default. At $10 per million input tokens and $50 per million output tokens, it is the most expensive model in the current PickAI roster, and its published input and output token-per-minute limits are only one quarter of the rates published for standard current Claude models. Fable can also decline classifier-flagged requests involving cybersecurity, biology or chemistry, model distillation, or frontier-AI development; supported Claude clients may then use safeguard fallback routing. Anthropic says client users are notified and API responses identify the serving model or refusal, so production integrations should surface that metadata and never assume every request was answered by Fable.

Ship Claude Sonnet 5 as the default Sonnet upgrade for production coding, agentic, and long-context workloads; budget for the September price step-up and reserve Opus for tasks that justify the higher tier.

Claude Sonnet 5 is the right default Sonnet model for most production teams that want stronger coding, agentic, and multi-step performance without moving every workload to Opus. It replaces Sonnet 4.6 in the current Claude lineup, carries stronger accepted-source HLE and SWE-Bench Pro evidence, and adds a 1M token context window that materially improves RAG, repository-scale coding, and document-heavy workflows. The migration still deserves an engineering review: the public row uses Anthropic vendor-reported benchmark evidence, the introductory API price ends after August 31, 2026, and plan-level availability or limits can vary by Claude surface. Treat it as the new default, test your real prompts and retrieval payloads, and route to Opus only when the task genuinely needs the higher-cost tier.