Hands-on with Claude Fable 5.2: 10 tasks burn 24% of an Ultra quota, front-end output stuns
数字生命卡兹克 · wechat · 2026-11-10
A blogger's overnight hands-on review of Claude Fable 5.2: 1M context, 128K output, 20% price cut, 30% faster, far fewer false safety refusals (55% less). He ran 10 single-prompt tasks on Xhigh — a fact-checked ChatGPT anniversary motion video, Three.js 3D scenes, and a 2.5-hour self-directed 3D rebuild of Apple's Jing'an store — burning 24% of his Ultra quota in 4 hours. His verdict: Fable 5.2 is best as a 'commander' model for wish-style programming, with Opus 5.5 handling front-end polish and the upcoming GPT-6.2 Sol handling algorithms and backend.
More from coding & agent
- Vera launches gyms, benchmarks and agents for cloud infrastructure operations — tianyin_xu · 2026-11-10
- Illos.dev sells $69 agent skills that draw on-style SVG illustrations — sull · 2026-11-10
- Wisdom: an open-source local ledger for Claude Code and Codex agent sessions — Fit-Fail-3369 · 2026-11-10
- Claude Code deletes session transcripts after 30 days by default — one config line keeps them — Kyrannio · 2026-11-10
- Morph launches Ultra: unlimited Kimi and DeepSeek for $200/month — zack_overflow · 2026-11-10
- AWS Nitro Sandboxes launch as lightweight VMs for hardware-isolated AI agents — tekbog · 2026-11-10