MineBench quirk: cheaper models Sonnet 5.5 and Sol 6.1 both attempt knight shadows, Opus doesn't
Separate_Tonight9533 · reddit · 2026-11-08
A Reddit user spotted a curious pattern on the MineBench graphics benchmark: both Sonnet 5.5 and Sol 6.1 — the cheaper-tier models — spontaneously try to add a shadow behind a knight figure, while Astra and Opus don't. Both execute shadows poorly, but the concept of light direction exists in the smaller models. The author wonders whether cheaper training paradoxically instills such understanding, or whether it's just coincidence.
More from Models
- Same model, same config: Reddit user documents rare episodes where agents act like a lousy support rep — No_Dragonfruit_1599 · 2026-11-08
- Unverified report: DeepSeek V4 Pro quietly updated, hard agent task pass rate jumps 16 points — teortaxesTex · 2026-11-08
- User frustrated as Anthropic monthly promo credits expire within 30 days — arthurcolle · 2026-11-08
- Dev calls out Anthropic hype: Boris Cherny claimed coding was 'solved' in February amid buggy Claude Code — kuza55 · 2026-11-08
- DeepSeek 1017 accused of wild benchmaxxing: great scores, not a frontier feel — teortaxesTex · 2026-11-08
- Why can LLMs solve PhD-level math but still lose at chess? Reddit asks — MesloubFC · 2026-11-08