Quality-adjusted cost rework of Artificial Analysis data shows cheap non-thinking models win
ResilientSpider · reddit · 2026-10-22
A Reddit user rebuilt the Artificial Analysis leaderboard around cost-per-task, time-per-task and the Intelligence Index, using a LiveBench-style "quality-adjusted cost" formula.
- The chart shows Luna and Haiku without thinking offer the best value: the extra intelligence of pricier models doesn't justify their cost and added latency.
- The author suggests an agentic, multi-turn automated approach could push efficiency even further.
Takeaway: for most tasks, small fast models beat flagship models on value.
More from Models
- unsloth ships GGUF quantizations of RecursiveSI/flint on Hugging Face — unsloth · 2026-10-22
- Generalist AI's GEN-1.5 puck-packing bot shows emergent physical reasoning and backtracking — chris_j_paxton · 2026-10-22
- User suspects OpenAI's "Approve for me" feature is behind recent token-draining spike — ssh4net · 2026-10-22
- Dev names Kimi K2 his favorite LLM: won't sycophantically agree, may outcode Sonnet 4 — gnukeith · 2026-10-22
- 31,000-question study finds LLMs across 24 languages are secretly obsessed with Japan — dioscuri · 2026-10-22
- Wiz says Gemini 4 Argon helped find decade-old critical vulnerability, coordinated fix across industry — HaydnBelfield · 2026-10-22