Kimi K3.1 Pro Costs ~20x DeepSeek V4.1 Flash Per Task, Blog Chart Slammed as Misleading
teortaxesTex · x · 2026-10-24
A viral critique flags Moonshot's Kimi blog Terminal-Bench 4.0 cost-vs-score scatter as visually misleading: DeepSeek-V4.1-Flash-1017 is shrunk to a tiny dot in the corner, obscuring that Kimi K3.1 Pro costs roughly 20x more per completed task ($4-5 at max tier vs under $1, with scores 65.7 vs 42.3).
teortaxesTex adds that the blog is internally inconsistent: the table cites 42.3 for DS-1017 from DeepSeek's own release, while the footnote claims all data come from tbench.ai except DeepSeek's, which come from Artificial Analysis.
Related event: Kimi K3.1 tops Terminal-Bench but costs far more than DeepSeek(3 posts)→
More from Models
- MazeBench Chart: Open-Weight Models Still Far Behind GPT-6 Astra (16% vs 1-2%) — patience_cave · 2026-10-25
- DeepSeek v4.1 Flash Makes Big MazeBench Leap Over Predecessor, Near Kimi's 3T Model — patience_cave · 2026-10-25
- Kimi K3.1 Pro Solves Multistep MazeBench Puzzle After 20 Tries, Only 5th Model Ever — patience_cave · 2026-10-25
- Flint ranked a top-3 daily coding driver, running Recursive_SI's local model with day-zero MLX support — altryne · 2026-10-25
- 177-case eval: cheapest LLM matched models costing 5–11x more for task parsing — danielbgolan · 2026-10-24
- Abacus AI to launch SuperComputer: text prompt to custom-trained LLM in an hour — bindureddy · 2026-10-24