UCSB AI Benchmarks

SciCode Leaderboard

Scientific coding ability on the SciCode benchmark: models solve research-level problems step by step, and each step is executed and checked against gold test cases.

Total Runs
30
Models
22
Providers
3
Last Run
Sep 16, 2026
Provider
Split
Metric
History
Showing 14 of 22 models — sorted by Main Resolve Rate (%) (higher is better)
#ModelRunsMain Resolve Rate (%)Sub-step Accuracy (%)Tokens / TaskOutput TokensTime / Task (min)StepsLast Run
🥇
GRITgpt-oss:20bvalidation
1
3.6K
53.3K
3
28 / 50Aug 5, 2026
🥈1
1.6K
24.3K
3
24 / 50Aug 5, 2026
🥉
GRITphi4:latestvalidation
1
2.1K
30.9K
4
22 / 50Aug 5, 2026
41
2.7K
178.5K
1
143 / 291Sep 15, 2026
53
2.8K
180.0K
3
118 / 291Sep 14, 2026
62
5.9K
383.7K
4
107 / 291Sep 15, 2026
72
3.5K
229.0K
2
100 / 291Sep 15, 2026
8
GRITllama3.1:8bvalidation
1
1.9K
28.1K
2
7 / 50Aug 5, 2026
91
4.5K
289.8K
2
103 / 291Sep 15, 2026
101
1.8K
116.9K
1
47 / 291Sep 15, 2026
111
1.4K
89.3K
1
67 / 291Sep 16, 2026
121
1.9K
123.3K
3
34 / 291Sep 14, 2026
132
1.2K
76.3K
2
24 / 291Sep 14, 2026
141
1.2K
18.1K
25
4 / 50Aug 5, 2026
Score bars show relative standing within the visible selection; token and time bars show magnitude. Each row reflects the most recent run per model, provider, and split. Click a score cell to expand run details.