ZOLT Model Test

48 AI models. The same 20 jobs. One strict ruler.

Run on the live ZOLT gateway · October 11, 2026 · 960 calls · Total measured cost: 41.91 credits

The question

Which AI model is cheap and good? Every model on ZOLT took the same 20 jobs — 5 writing, 5 questions, 5 code, 5 logic — at temperature 0. Every answer was scored by rule, no human judging, and the code answers were actually run and tested. Same jobs, same ruler, no favorites.

Winners

Perfect score — 20/20

MiMo V2.6 Flash · GLM 5.3 Flash · GLM 5.3 FlashX

Cheapest good model — all 4 categories

MiMo V2.6 Flash and GLM 5.3 Flash (5/5 in Writing, Questions, Code, and Logic, at about 0.001 credits per call)

Cheapest strong all-rounder

GPT-6 Luna — 18/20 with perfect Code and Logic, about 0.001 credits per call

Most expensive per call

Claude Sonnet 5 — 0.356 credits per call, scored 17/20

What the test showed

Price is not quality. Three near-free models scored a perfect 20/20 while famous flagships costing 100x more per call landed at 18/20 — and the most expensive model on the board scored 17/20.
Two traps set the ceiling. One question demanded the answer written with its unit ("150 km") — 41 of 48 models failed it. One writing job required a thank-you line — 39 failed that. The test measures following instructions exactly, which is where real AI work succeeds or fails.
Code separated the field. Every model scored 5/5 or 4/5 on code except GPT-5.4 and GPT-5.4 Mini, whose answers came back un-runnable (1/5) and finished last at 10/20.
Speed varies wildly. The fastest models answer a job in about 3.5 seconds (median). GPT-5.6 took about 47 seconds — 13x slower — for a middle-of-the-pack 17/20.

Full results

#	Model	Score	Writing	Questions	Code	Logic	Median time	Credits / call
1	MiMo v2.6 Flash	20/20	5/5	5/5	5/5	5/5	4.8s	0.001
2	GLM 5.3 Flash	20/20	5/5	5/5	5/5	5/5	4.3s	0.001
3	GLM 5.3 FlashX	20/20	5/5	5/5	5/5	5/5	4.8s	0.005
4	HY4 Preview	19/20	5/5	4/5	5/5	5/5	4.5s	0.006
5	Kimi K2.8	19/20	5/5	4/5	5/5	5/5	12.9s	0.012
6	GPT-6 Luna	18/20	4/5	4/5	5/5	5/5	3.5s	0.001
7	DeepSeek V4.1 Flash	18/20	4/5	4/5	5/5	5/5	4.4s	0.003
8	Qwen 3.8 Omni Flash	18/20	4/5	4/5	5/5	5/5	4.5s	0.003
9	Grok 4.6	18/20	4/5	4/5	5/5	5/5	4.3s	0.004
10	Gemini 3.7 Flash	18/20	4/5	4/5	5/5	5/5	3.9s	0.005
11	GLM 5.1	18/20	4/5	4/5	5/5	5/5	3.6s	0.006
12	GPT-6 Sol	18/20	4/5	4/5	5/5	5/5	5.0s	0.007
13	MiniMax M3	18/20	4/5	4/5	5/5	5/5	4.2s	0.008
14	DeepSeek V4 Pro	18/20	4/5	4/5	5/5	5/5	4.1s	0.009
15	Claude Haiku 4.5	18/20	4/5	4/5	5/5	5/5	5.0s	0.010
16	Qwen 3.7 Max	18/20	4/5	4/5	5/5	5/5	6.1s	0.019
17	GPT-5.5	18/20	4/5	4/5	5/5	5/5	4.0s	0.019
18	Claude Opus 5	18/20	4/5	4/5	5/5	5/5	3.9s	0.019
19	Claude Opus 4.8	18/20	4/5	4/5	5/5	5/5	3.7s	0.020
20	Qwen 3.8 Max	18/20	4/5	4/5	5/5	5/5	4.5s	0.020
21	GLM 5.2	18/20	4/5	4/5	5/5	5/5	4.8s	0.020
22	MiniMax M2.7	18/20	4/5	4/5	5/5	5/5	10.3s	0.022
23	Gemini 3.8 Flash	18/20	4/5	4/5	5/5	5/5	6.4s	0.024
24	Claude Opus 4.7	18/20	4/5	4/5	5/5	5/5	4.5s	0.024
25	Qwen 3.8 Max Preview	18/20	4/5	4/5	5/5	5/5	6.0s	0.026
26	GLM 5.3	18/20	4/5	4/5	5/5	5/5	5.8s	0.031
27	GPT-5.6 Terra	18/20	4/5	4/5	5/5	5/5	5.4s	0.033
28	Grok 4.7	18/20	4/5	4/5	5/5	5/5	4.5s	0.034
29	Kimi K3	18/20	4/5	4/5	5/5	5/5	7.9s	0.041
30	Grok 4.5	18/20	4/5	4/5	5/5	5/5	5.3s	0.047
31	GPT-6.1 Sol	18/20	4/5	4/5	5/5	5/5	5.0s	0.067
32	Gemini 3.1 Pro Preview	18/20	4/5	4/5	5/5	5/5	5.8s	0.068
33	Claude Sonnet 4.6	18/20	4/5	4/5	5/5	5/5	4.7s	0.090
34	Claude Sonnet 5.5	18/20	4/5	4/5	5/5	5/5	6.0s	0.106
35	GPT-5.6 Sol	18/20	4/5	4/5	5/5	5/5	13.5s	0.111
36	GPT-6 Astra	18/20	4/5	4/5	5/5	5/5	5.7s	0.148
37	Claude Fable 5	18/20	4/5	4/5	5/5	5/5	6.3s	0.170
38	Claude Opus 5.5	18/20	4/5	4/5	5/5	5/5	5.0s	0.212
39	Kimi K2.6	17/20	3/5	4/5	5/5	5/5	4.8s	0.012
40	GPT-5.6	17/20	3/5	4/5	5/5	5/5	46.8s	0.066
41	Gemini 3 Pro Preview	17/20	3/5	4/5	5/5	5/5	7.1s	0.082
42	Claude Sonnet 5	17/20	4/5	4/5	5/5	4/5	6.3s	0.356
43	Kimi K2.7 Code	16/20	3/5	4/5	4/5	5/5	4.7s	0.036
44	Grok 4.3	15/20	4/5	4/5	4/5	3/5	3.7s	0.002
45	HY3	15/20	2/5	4/5	4/5	5/5	5.6s	0.003
46	MiMo v2.5	14/20	3/5	4/5	4/5	3/5	7.2s	0.003
47	GPT-5.4 Mini	10/20	3/5	4/5	1/5	2/5	5.7s	0.006
48	GPT-5.4	10/20	3/5	3/5	1/5	3/5	4.3s	0.019

How it was scored

Writing jobs were checked for required points, word limits, and banned words. Questions each have one right answer, checked exactly (including the unit where the job asked for one). Code answers were executed in a locked-down sandbox and tested against the expected behavior. Logic answers were matched exactly. One retry pass was run for 6 calls that errored on the first pass (merchant-side failures, including all 5 of one model's writing calls); the retried answers are the ones scored above. Costs are computed from each call's real token usage at ZOLT's live credit rates and cross-checked against the gateway balance: 41.91 credits burned in total, measured. Full raw results (every answer, every score) are available on request.