MMODELYST
Signals · novel-idea finder

Discover

Benchmark correlation

Pearson r across models scored on both benchmarks. Amber = redundant (measures the same thing); blue = independent signal. Pick benchmarks that disagree to cover more ground.

aa-lcr
aime
bbh
gpqa
gsm8k
hle
humaneval
ifbench
ifeval
livecodebench
math
math-lvl5
mmlu
mmlu-pro
mmmu
musr
scicode
swe-bench
tau2-bench
terminal-bench-hard
aa-lcr·0.70.80.70.80.70.40.60.80.80.80.80.8
aime0.7·0.90.70.70.90.70.70.70.70.60.6
bbh·0.70.70.6-0.10.20.40.90.90.5
gpqa0.80.90.7·0.80.70.80.8-0.20.90.80.30.90.90.80.60.90.80.70.8
gsm8k0.70.8·0.90.80.50.80.80.90.6
hle0.70.70.7·0.20.80.70.40.50.50.70.70.80.70.9
humaneval0.60.80.90.2·0.20.90.80.80.90.90.80.90.50.9
ifbench0.80.70.80.80.2·0.80.50.80.70.70.8-0.00.80.8
ifeval-0.1-0.20.80.9·-0.2-0.00.9-0.2-0.3
livecodebench0.70.90.90.70.80.8·0.80.80.80.80.80.30.70.7
math0.40.70.20.80.50.40.80.5-0.20.8·0.60.70.90.80.40.70.70.50.5
math-lvl50.40.30.80.9-0.00.6·0.70.60.3
mmlu0.90.90.80.50.90.80.90.80.70.7·0.90.90.80.8
mmlu-pro0.60.70.90.90.90.50.80.7-0.20.80.90.60.9·0.90.60.90.80.60.6
mmmu0.80.70.80.70.90.70.80.80.90.9·0.90.50.70.8
musr0.50.60.60.5-0.30.40.30.80.6·
scicode0.80.70.90.70.90.80.80.70.80.90.9·0.90.70.8
swe-bench0.80.80.8-0.00.30.70.80.50.9·0.61.0
tau2-bench0.80.60.70.70.80.70.50.60.70.70.6·0.8
terminal-bench-hard0.80.60.80.90.80.70.50.60.80.81.00.8·
Most redundant pairs
gpqa × livecodebenchr=0.93 · n=276
aime × livecodebenchr=0.91 · n=207
gpqa × scicoder=0.91 · n=401
ifeval × musrr=-0.30 · n=1480
ifeval × mmlu-pror=-0.22 · n=1480
gpqa × ifevalr=-0.20 · n=1482
Capability coverage — white space

How well each capability is measured (scored model-benchmark pairs). Thin bars = under-explored — opportunity.

Multilingual
0 bench · 0 scores
Safety & Alignment
0 bench · 0 scores
Vision & Multimodal
1 bench · 8 scores
Instruction Following
1 bench · 330 scores
Long Context
1 bench · 339 scores
Agents
3 bench · 645 scores
Code
4 bench · 697 scores
Language & Instruction
1 bench · 1488 scores
Knowledge & QA
2 bench · 1775 scores
Math
4 bench · 1891 scores
Reasoning
5 bench · 5273 scores
Efficiency frontier

Best average benchmark score achievable at each model size (Pareto frontier) — models that punch above their weight.

Price vs performance — best value

Average benchmark score per dollar of output (per 1M tokens), among models with published pricing. Higher value = more quality for your money. Prices are vendor-reported — verify before relying.

ModelAvg$/M in·outValue
1Gemma 3n E4B InstructGoogle21$0.02 · $0.04528
2Qwen3.5 9BAlibaba51$0.1 · $0.15341
4Qwen3.5 4BAlibaba46$0.03 · $0.15304
5gpt-oss-20bopenai52$? · $0.2261
8DeepSeek V4 FlashDeepSeek63$0.14 · $0.28224
9MiMo-V2-FlashXiaomi66$0.1 · $0.3221
10Ministral 3 3BMistral22$0.1 · $0.1218
11MiMo-V2.5Xiaomi59$0.14 · $0.28212
12Llama 3.1 Instruct 8BMeta20$0.075 · $0.095212
Model release velocity

Models in the catalog by release quarter (last 3 years).

23 Q3
23 Q4
24 Q1
24 Q2
24 Q3
24 Q4
25 Q1
25 Q2
25 Q3
25 Q4
26 Q1
26 Q2
26 Q3