Benchmark

Aggregate win rates from blinded side-by-side triage comparisons. User-judged preferences, not objective correctness.

8 runs · 1 voter · updated just now

1GPT-5.4 Mini

67%win rate

4 wins6 runs4,345ms0.33¢90 conf.Default

2GLM 5.2

43%win rate

3 wins7 runs14,286ms0.88¢75 conf.Default

3GPT-5.4 Nano

33%win rate

1 wins3 runs4,629ms0.12¢86 conf.Default

Standings

Ranked by win rate. Wilson 95% confidence interval shown.

#	Model	Wins	Runs	Win rate	95% CI	Avg conf.	Avg latency	Avg cost	Thinking	Blind / Reveal
1	GPT-5.4 Mini OpenAI	4	6	67%	30–90%	90	3,987ms	0.33¢	Default	4 / 0
2	GLM 5.2 Z.ai	3	7	43%	16–75%	75	16,845ms	0.88¢	Default	3 / 0
3	GPT-5.4 Nanolow sample OpenAI	1	3	33%	6–79%	86	3,810ms	0.12¢	Default	1 / 0
4	DeepSeek V4 Flashlow sample DeepSeek	0	2	0%	0–66%	—	—	0.10¢	Default	0 / 0
5	DeepSeek V4 Prolow sample DeepSeek	0	1	0%	0–79%	—	—	1.11¢	Default	0 / 0
6	Gemma 4 31Blow sample Google	0	2	0%	0–66%	—	—	0.04¢	Default	0 / 0
7	MiniMax M3low sample MiniMax	0	1	0%	0–79%	—	—	0.12¢	Default	0 / 0
8	Kimi K2.6low sample Moonshot AI	0	1	0%	0–79%	—	—	1.76¢	Default	0 / 0
9	GPT-5.5low sample OpenAI	0	1	0%	0–79%	—	—	2.04¢	Default	0 / 0
10	Grok 4.3low sample xAI	0	4	0%	0–49%	—	—	1.20¢	Default	0 / 0
11	MiMo M2.5low sample Xiaomi	0	3	0%	0–56%	—	—	0.72¢	Default	0 / 0

Speed vs accuracy

Each dot is a model. Top-left is fastest and most accurate. Hover for detail. Outlined dots are low sample.

Win rate

Fast & accurateSlow & accurateFast & less accurateSlow & less accurate

GPT-5.4 Mini

GLM 5.2

GPT-5.4 Nano

0ms

11,654ms

48,307ms

Avg latency

Methodology

Wins are user-judged preferences on blinded side-by-side runs — not objective correctness. Wilson 95% confidence intervals are shown on every win rate; models with fewer than 5 appearances are flagged as low sample. Average confidence is self-reported by the winning model. Synthetic cases only. Research demo, not for clinical use.