Fable 5.2 Reviews Itself From 160 User Reports: 5 Rules for Using It Cheaply
henrik392 · reddit · 2026-11-10
The author had Fable 5.2 review 160 day-one user reports and produce an Anthropic-style self-review video. Key findings: it can work autonomously for days from vague prompts (Matthew Berman got a playable shooter level, 3D Tokyo); large builds cost an estimated $11k–15k at API prices; Theo's logs show a $200 plan buys 1/5 the value of Opus 5.5; ordinary coding ties Opus 5.5 (26/30 both) at $1.78 vs $1.00 per task; but it pulls ahead on big jobs, fixing 58 of 105 planted bugs vs Opus 5.5's 41.7 on Bug Hunt Bench, and grades its own work poorly (a 182-page paper). Five rules: start with Opus 5.5, give Fable the vague big job, let it direct while Opus codes, keep effort at medium/high ($9.08/task at high vs $26.54 at max), never let it grade itself. The video was built with 6 Sonnet subagents researching sources plus HTML/CSS frames rendered via headless Chrome and ffmpeg.
More from coding & agent
- Ex-Zalando AI lead's thread: agentic AI changes what a company is, not just add-ons — ilparone · 2026-11-10
- AIMUG meetup: with agents you're the manager—define work, context and tools — colinmcnamara · 2026-11-10
- Fable 5.2 built an interactive 3D globe of Earth's history from one prompt, ~$550 in tokens — Chris_Brannigan · 2026-11-10
- Adobe ships MCP support across its entire Creative Cloud stack, from Photoshop to Firefly — LinusEkenstam · 2026-11-10
- Hugging Face Transformers merges decision models primitive: Jev, KEV, CLEF, Decider — multimodalart · 2026-11-10
- Microsoft's LongCapHarness boosts Llama3-8B long-context by 64 points with zero training — omarsar0 · 2026-11-10