MoonMapBench: Opus 5.5 Off by 1,192 m Predicting Lunar Elevations
maksym_andr · x · 2026-10-13
fullrank introduced MoonMapBench, an "unsaturated frontier eval" that scales to any planet or moon with a known elevation profile: AI must predict elevations at 16,200 lunar coordinates. Claude Opus 5.5 averages 1,192 m (3,911 ft) off; GPT 6 Astra is worse at 1,490 m (4,888 ft). The author quips that frontier evals should measure things beyond Earth, and the benchmark can extend to any celestial body, making it hard to saturate.
More from Fun
- Email is still hard in the age of AI, and everyone loves a test email — fhinkel · 2026-10-13
- Musk bragged about his Grok Lego Star Wars haul, then X users punked him — Wired AI · 2026-10-13
- One-liner: Anthropic remains the only company that soulmaxxes its models — cto_junior · 2026-10-13
- Tech people will never have taste? The AI 'soulfulness' discourse kicks off — basedjensen · 2026-10-13
- OpenAI 'dots' hardware spotted in the wild, stuck at UK customs — SamuelAlbanie · 2026-10-13
- Strawberry set the tone: every model release since is now a fruit — Kyrannio · 2026-10-13