Towards AITowards AIToneBench

GPT-6 Sol (max) vs Kimi K3

Head-to-head on the Towards AI writing benchmark: same 10 real YouTube scripts, five runs each, scored blind by a three-family judge panel. Kimi K3 leads overall, 88.5 to 88.0. Their per-model Elo intervals overlap on this board.

GPT-6 Sol (max)
#31 Elo 2166 · 88.0/100
Kimi K3
#21 Elo 2217 · 88.5/100
Cost / script
$0.112 vs $0.260
Human baseline
90.9 GPT-6 Sol (max) falls below it · Kimi K3 falls below it

The verdict

This one is closer than the ranks suggest. Kimi K3 sits at #21 with 2217.0 Elo and 88.48 overall; GPT-6 Sol (max) is at #31 with 2165.6 Elo and 87.95 overall. Their 95% Elo intervals overlap (2179.1 to 2248.4 against 2127.0 to 2200.7), so do not read much into the exact order. Where Kimi K3 pulls ahead: Tone & Voice Match (88.39 vs 85.84) and Hook Strength (89.24 vs 86.91). GPT-6 Sol (max) still wins on Visual Cue Quality (89.15 vs 85.14) and Slop Score (EQ-Bench + ours) (94.08 vs 91.23), so it is not a clean sweep. The trade is price: GPT-6 Sol (max) costs about $0.112 per article against $0.260 for Kimi K3, so you pay for the higher rank. Kimi K3 publishes open weights; GPT-6 Sol (max) does not. Default to Kimi K3. Reach for GPT-6 Sol (max) when visual cues, slop score, and the lower price matter more to you than the headline rank.

Pick Kimi K3 for the stronger board result, tone, and the hook.
Pick GPT-6 Sol (max) for visual cues, slop score, and the lower price.

Reading the Elo intervals. Each range estimates uncertainty around one model's Elo. We repeatedly sample from every model's recorded run scores on each task and recalculate the full ranking. Comparing these separate ranges does not test the Elo gap between two models. The tasks stay fixed, so the ranges do not measure how the ranking would change on new tasks.

Metric by metric

Blue bars: GPT-6 Sol (max). Orange bars: Kimi K3. Same 0–100 scale; the bold bar wins that metric.

Tone & Voice Match19% weight
85.8
88.4
Writing Craft & Clarity13% weight
87.8
89.0
Substance, Accuracy & Value15% weight
88.9
87.6
Continuity & Emotion14% weight
86.4
87.6
YouTube Best Practices12% weight
87.9
88.4
Hook Strength10% weight
86.9
89.2
Length Adherence8% weight
91.2
90.1
Slop Score (EQ-Bench + ours)5% weight
94.1
91.2
Visual Cue Quality4% weight
89.2
85.1

Everything else that differs

GPT-6 Sol (max)Kimi K3
Overall / 10088.088.5
Writing Elo21662217
Run-to-run spread (± overall std)1.3301.260
Cost per script (USD)0.1120.260
Avg latency (s)171.0234.3
Open weightsNoYes

Full scorecards: GPT-6 Sol (max) · Kimi K3. How scoring works: methodology.

← All comparisons