Night train benchmark: GPT-6 Luna beats Haiku 5.5 — 5x faster at a third of the cost
RexDouglass · x · 2026-10-08
Developer EnvolDev ran his self-styled "night train benchmark" comparing Haiku 5.5 against GPT-6 Luna: Luna finished in 54.6 seconds for under $0.01, while Haiku 5.5 took 5 minutes 12 seconds at $0.03. Luna was over 5x faster, less than a third of the cost, and produced a noticeably better result. The author concludes Haiku 5.5 "completely failed" this test. Note: a one-person, non-standard benchmark.
More from Models
- Polymarket launches betting market on when Google's Gemini Argon will be released — Polymarket · 2026-10-08
- Local Models Are Dead, Says Blogger: Microsoft's Hybrid Intelligence Is the Real Road — TheOyinbooke · 2026-10-08
- GPT-6 reportedly identifies itself as "GPT 5.6 Sol" across multiple sessions — Super_Pattern6594 · 2026-10-08
- Sebastian Raschka traces text classification from bag-of-words to the viral Jev decision model — AxSaucedo · 2026-10-08
- Why assume open-weight model providers will stay open forever? — Atlan_ · 2026-10-08
- Qwen Flash ultra-low-bit quants compared: Q2_0 steadier, IQ2_XS keeps knowledge but loops — dampflokfreund · 2026-10-08