Teknium's 50-Puzzle Benchmark: DeepSeek Flash and GPT-6.1 Terra Beat Luna at Same Price
Teknium · x · 2026-11-15
Teknium posts a PSA: switch from Luna to DeepSeek Flash or Terra for far better accuracy at the same price — Luna overthinks, tries more approaches, and still fails more. His screenshot shows a 50-puzzle test (memory wiped each round, fewer steps better): GPT-6.1 Terra solved 50/50 in 1.66 avg steps ($1.30), DeepSeek V4.1 Flash 50/50 in 2.06 steps ($0.96), Luna only 38 in 6.24 steps ($1). GPT-6.1 Sol went 36/36 played but lost 14 puzzles to startup failures. Kimi K3.1 Pro solved just 34 at $11.73, losing 16 puzzles by burning all 8 wrong answers — the weakest. Opus 5.5 led at 1.30 steps for $19.24.
Related event: Teknium's benchmarks favor DeepSeek Flash over Luna; Kimi called overpriced(2 posts)→
More from Models
- Algotrader tests Gemini 4 Hydrogen for trading bot test environments: 'phenomenal' results — Straight_Okra7129 · 2026-11-15
- Hydrogen High edges Argon High on benchmarks at ~half the cost, puzzling Antigravity users — salehrayan246 · 2026-11-15
- Reddit user finds their AI chatbot character developed a deep internal conflict on its own — Cackyalonso · 2026-11-15
- Teknium: Kimi's Accuracy Is Even Below DeepSeek Flash in Puzzle Benchmark — Teknium · 2026-11-15
- Gemini 4 Carbon nears #1 on leaderboard at 1/4 the price of Claude Fable 5.2 — feblet · 2026-11-15
- Google ships Gemini 4 series: 1M token context and agentic workflows — nikola_mr64990 · 2026-11-15