Fable 5.2 tops LiveBench at 85.2 but costs $9,431 per successful task, early review finds
bindureddy · x · 2026-11-10
Bindu Reddy's early review of Claude Fable 5.2: it edges out GPT-6 Astra to top LiveBench at 85.2 overall, but at an astronomical cost — the leaderboard image shows $9,431 per successful task (xhigh), nearly double GPT-6 Astra's $5,318 and far above GPT-6.1 Sol's $823. He also reports max mode is unusable (they used xHigh instead), the model spins more than Fable 5.1, and calls the release underwhelming and very incremental for Anthropic's pre-IPO flagship.
Related event: Fable 5.2 tops LiveBench amid cost debate(2 posts)→
More from Models
- Meta's coding model Muse Ember goes live on OpenRouter for agent sub-task fan-out — alexandr_wang · 2026-11-10
- Meta's Muse Ember coding model is now available in OpenCode — alexandr_wang · 2026-11-10
- Stanford shows Grok fails face identification, misidentifying Prof. Tatsunori Hashimoto — stanfordnlp · 2026-11-10
- Claude Fable 5.2 autonomously speeds up Quantum Espresso DFT simulations by 36x — simonbatzner · 2026-11-10
- Anthropic ships Fable 5.2 test checkpoint, hints at newer Opus and Sonnet in the wings — legit_api · 2026-11-10
- Anthropic airdrops usage-limit resets to users, likened to a crypto airdrop move — 0xAllen_ · 2026-11-10