The supplied evidence identifies the target as Claude Fable 5, rather than the catalog label “Claude Fable.” It has unusually strong capability evidence: Anthropic reports 95.0% on SWE-bench Verified, DeepSWE lists 70%±4% (third behind Claude Opus 5 and GPT-5.6 Sol), and Terminal-Bench lists Claude Code with Fable 5 first at 83.8%±1.2%. These results support a coding score just below GPT-5.5/GPT-5.6 Sol’s 96/95 anchors, while its 1M context and 128k maximum output support high technical capability. Arena’s rank-one 1507±6 is supportive preference evidence, not independent proof of reasoning quality. Fable 5 compares favorably with Anthropic’s published Opus 4.8 anchor on context, documented agent performance, and apparent throughput; however, its coding score remains slightly lower because the strongest SWE-bench result is provider-reported and DeepSWE trails the stated leaders. Its $10/MTok input and $50/MTok output price is materially less costly than a typical premium frontier tier, but output pricing remains expensive relative to efficient models such as Gemini Flash or Kimi. Artificial Analysis reports 62 output tokens/s, a solid but not exceptional speed signal. Developer experience is strong through the Claude API, explicit model ID, caching prices, and long-context limits. Pricing documentation is detailed, but availability is materially unclear: the supplied product-page excerpt says unavailable while API documentation says general availability began June 9, 2026. Multimodal capability is not sufficiently specified in the provided excerpts, so it should not receive frontier-leading credit. The evaluation should remain conditional on confirming that the official pages and current API access are live.
