Kimi K3 (thinking) vs DeepSeek V4 Flash 0731
Should you write with Kimi K3 (thinking) or DeepSeek V4 Flash 0731? On this page the only thing that changes is the model. Both wrote their own version of the scripts behind 10 of my What's AI videos, five drafts per script, and three judges from three model families scored every draft blind, with my finished script as the reference. Kimi K3 (thinking) leads overall, 87.2 to 84.4.
The verdict
Kimi K3 (thinking) is the better writer, and it's also the quicker one, which is not what you'd expect from a thinking setting. It ranks #41 to #59 and wins all nine metrics, with voice and continuity ahead by roughly three and four points. It turns a script around in a little over a minute, against about three for DeepSeek V4 Flash 0731. One tip if you choose Kimi: the default Kimi K3, without thinking, ranks higher still. DeepSeek's case is cost, about half a cent a script against $0.095, with open weights on both sides, so it wins when you need a mountain of drafts.
Pick Kimi K3 (thinking) when you want better, faster scripts from an open-weights model.
Pick DeepSeek V4 Flash 0731 when half a cent per script is the number that matters.
How to read these numbers
Overall vs Elo. Overall is the weighted rubric score. Elo compares each model's five-draft average with every other model's, script by script, and counts gaps inside run-to-run noise as draws, so it rewards winning often and can disagree with overall. My own scripts score 90.9 on the same rubric, not a literal 100. Right now 2 of 167 ranked configurations reach it. A model above that line writes at that level on this rubric. That's not the same as beating my script line for line.
Reading the Elo intervals. Each 95% range comes from resampling every model's recorded run scores on each task and re-ranking. Comparing these separate ranges does not test the Elo gap between two models: far apart is strong evidence, and overlap isn't a tie. The tasks stay fixed, so the ranges say nothing about new tasks (the long version).
Metric by metric: where the gap comes from
Blue is Kimi K3 (thinking), orange is DeepSeek V4 Flash 0731, and the solid bar takes the metric. The weight under each name is how much it counts toward the overall. Weighted by how much each metric counts, Tone & Voice Match does the most to put Kimi K3 (thinking) ahead: it leads there by 3.2 points. If you care about one thing, say voice or length, go straight to that row.
Cost, speed, and consistency
Quality is half the decision. The other half is what each script costs you, how long you wait, and how often you get a bad draft. The lead has a price: Kimi K3 (thinking) costs about 19x as much per script as DeepSeek V4 Flash 0731.
| Kimi K3 (thinking) | DeepSeek V4 Flash 0731 | |
|---|---|---|
| Overall / 100 | 87.2 | 84.4 |
| Writing Elo | 2098 | 1866 |
| Score spread across drafts (± overall std) | 1.550 | 3.360 |
| Cost per script (USD) | 0.095 | 0.005 |
| Avg time per script (min) | 1.2 | 3.1 |
| Open weights | Yes | Yes |
Green wins the row, and for spread, cost and time, lower wins. Cost is one script at recorded list prices, uncached and without judging, so your real bill will differ. Spread is how much the overall score moves across a model's drafts, and time is the average time to get one accepted script, retries included.
Want the script-by-script detail? Full scorecards: Kimi K3 (thinking) · DeepSeek V4 Flash 0731. The whole scoring pipeline is on the methodology page.
← All comparisons