Skip to content

Benchmark

hal-scicode

Coding benchmarks compare models on programming tasks. Scores are percentages.

Models measured
11
Scores recorded
15
Sources
1
IQ weight
0.53

Score distribution

4.6%
Mean
4.6%
Median
3%
Std dev
0%
10th percentile
9.2%
90th percentile
0%
Min
0%9.2%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.53 for chat models, measured over 11 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#Model
1o4 Minilow· HAL (Princeton)9.2%
2OpenAI o3medium· HAL (Princeton)9.2%
3Claude Opus 4.1· HAL (Princeton)7.7%
4Claude Opus 4.1high· HAL (Princeton)6.9%
5GPT 4.1· HAL (Princeton)6.2%
6GPT 5medium· HAL (Princeton)6.2%
7o4 Minihigh· HAL (Princeton)6.2%
8Claude Sonnet 3.7high· HAL (Princeton)4.6%
9Claude Sonnet 3.7· HAL (Princeton)3.1%
10Claude Sonnet 4.5· HAL (Princeton)3.1%
11DeepSeek V3· HAL (Princeton)3.1%
12Claude Sonnet 4.5high· HAL (Princeton)1.5%
13Gemini 2.0 Flash· HAL (Princeton)1.5%
14Claude Haiku 4.5· HAL (Princeton)0%
15DeepSeek R1· HAL (Princeton)0%

Tokens the model generated per task, as measured by the source. Comparable within this benchmark only.

Data credits: HAL (Princeton) — cite the HAL paper