Claude Fable 5.2 retakes world #1 as Anthropic IPO roadshow begins — 20% price cut, costlier per task

Latent Space · rss · 2026-11-10

Latent Space's weekly roundup covers Claude Fable 5.2 / Mythos 5.2, launched as Anthropic's IPO roadshow begins. Key facts:

Claims & pricing: 'a step above Opus 5.5 in almost all capabilities'; $8/$40 per 1M tokens (20% below 5.1); 1M context; five effort levels, Medium default. But it's token-hungry: GDPx cost per task rose from $67.16 to $88.15 as the model writes hundreds of test cases and spawns subagents to verify. At Max effort, $26.5/task — 2.6× Opus 5.5.
Evals: #1 on Artificial Analysis Index (65 vs Gemini 4 Carbon 60, Opus 5.5 59); records on HLE-Diamond (80%), ARC-AGI-3 adapter (97.7%), Epoch index (174). Weaker on Terminal-Bench 4.0 (60%, behind Kimi K3.1 Pro's 67%).
Official demos: 1,000 agents formalized the Fermat Last Theorem proof in 30 hours; built a bootable OS from an empty folder over 8 days with kernel and drivers.
Restrictions: subscription resets scaled by tenure, Fable capped at 50% of weekly limits, mandatory data retention by default for safety classifiers. Day-one availability in Cursor, Copilot/VS Code, Perplexity; system-card wording implies a shared base model for Fable/Mythos.

More from Venture