Anthropic's Claude Fable 5, the first generally available Mythos-class model, hits 80% on SWEBench Pro, outperforming Opus 4.8, GPT-4.5, and Gemini 3.1 Pro. It costs $10 per million input tokens and $50 per million output tokens, a new pricing tier above Opus, and consumes tokens at roughly twice the rate of comparable models. Reviewer Claire ran it through product specs, agent workflows, vision tasks, and multi-agent orchestration to find out where that price is justified.
The model's strengths are narrow and specific. Vision tasks, particularly PDF parsing and document formatting, are genuinely ahead of Opus 4.8. Multi-agent orchestration is technically functional. But the weaknesses are hard to ignore: PRDs and spec writing produce dense, unnavigable blocks of text. One-shot UI design output was described as fundamentally bad, gray and red outlines with no visual judgment. The model interprets 'minimal MVP' so literally that the output delivers almost no customer value. Fable 5 behaves like a seasoned engineer who investigates every corner before shipping anything, which is the wrong tool when speed matters.
The safety architecture is worth reading about in detail. Fable 5 does not hard-block restricted queries in cybersecurity, biology, chemistry, and distillation. Instead it falls back to Opus 4.8, and Anthropic reports 95% of sessions never trigger that fallback. A 30-day data retention policy exists solely for misuse detection. The full review maps exactly which task types justify the cost and which do not, a practical decision framework for teams deciding where to route workloads.
[READ ORIGINAL →]