skynet · reasoning
SKYNET

minimax-m3

Herman's daily-driver skynet reasoning model

⚖ COMPARE MODELS
VALS INDEX
59.3% +/- 9.5
Accuracy
VALS INDEX
16.8s
Latency (median)
$VALS INDEX
$0.2393
Cost / Test (1P-est)
SCORE STATUSscored
TEST SETS13
RECORDED TRIALS811
CUSTOM PASS70%
DEEP VALIDATION · 2026-08-20

Dual-judge family campaign

FULL REPORT →
CONSENSUS MEAN3.922 / 5
BOTH-JUDGE PASS71.1%
CROSS-JUDGED45 / 45
MEDIAN WALL17.70s
GPQA_DIAMONDserialized verification sample
62.5%
40/40 responses
IFEVALserialized verification sample
62.5%
40/40 responses
MMLU_PROserialized verification sample
67.5%
40/40 responses

Industry benchmarks

Accuracyattempts · responded
GPQA_DIAMOND GPQA Diamond
52.4%
231 attempts · 201 responded
HUMANEVAL_PLUS HumanEval+
73.3%
120 attempts · 103 responded
IFEVAL IFEval
54.6%
160 attempts · 122 responded
MMLU_PRO MMLU-Pro
57.0%
200 attempts · 183 responded
Industry benchmark Bars = accuracy %, black cap = that score

Custom 9-task suite

Mean /5n=trials
CUSTOM json_strict
5.00
11 trials
CUSTOM reasoning_multistep
5.00
11 trials
CUSTOM code_gen_long
4.82
11 trials
CUSTOM agentic_prompt
4.64
11 trials
CUSTOM summarize
4.36
11 trials
CUSTOM agentic_tool_use
4.33
9 trials
CUSTOM refactor_existing_code
3.78
9 trials
CUSTOM code_debug
3.12
16 trials
CUSTOM creative_write -
0.82
11 trials
Custom suite · scored 0–5 by consensus (m3 + Fable-5) ★ = task scored a 5 at least once · sorted by mean score

Cross-judge validation

m3 vs Fable-5 Δ

No judge bias detected on this model — m3 and Fable-5 agree across all tasks.

Per-task distribution

Score

The accumulated custom suite below uses minimax-m3 + claude-fable-5 consensus. The dedicated 2026-08-20 campaign above is supplemental and uses minimax-m3 + a blind grok-4.6 re-judge on every successful response. Industry suites use deterministic family-specific graders. See methodology and the campaign report.