Scientific Code Benchmark (SciCode)
SciCode evaluates language models on generating code for realistic scientific research problems across 16 subfields of physics, math, chemistry, biology, and material science. Problems decompose into 338 subproblems requiring domain knowledge recall, scientific reasoning, and precise code synthesis. Based on real scripts from published research.
Top models on SciCode — October 29, 2026
As of October 29, 2026, Sakana Fugu leads the SciCode leaderboard with 60.1% , followed by Step 5 Preview (58.9%) and Sakana Fugu-Ultra (58.7%).
Sakana Fugu
Sakana AI
Step 5 Preview
StepFun
Sakana Fugu-Ultra
Sakana AI
35 modelsCoding10% of Coding reference weightRefreshingUpdated October 29, 2026
Leaderboard (35 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Sakana FuguSakana AI | 60.1% | Not reported | Closed |
| 2 | Step 5 PreviewStepFun | 58.9% | Not reported | Open |
| 3 | Sakana Fugu-UltraSakana AI | 58.7% | Not reported | Closed |
| 4 | Qwen3.7 MaxAlibaba | 53.5% | Not reported | Closed |
| 5 | Gemini 3.5 FlashGoogle | 53.1% | Not reported | Closed |
| 6 | Kimi K2.6Moonshot AI | 52.2% | Not reported | Open |
| 7 | Qwen3.7 PlusAlibaba | 51.3% | Not reported | Closed |
| 8 | Beam 1.0 501B-A23BReflection AI | 49.2% | Not reported | Open |
| 9 | Inkling-SmallThinking Machines Lab | 48.7% | Not reported | Open |
| 10 | Kimi K2.5Moonshot AI | 48.7% | Not reported | Open |
| 11 | Beam 1.0 501B-A23B FP8Reflection AI | 48.3% | Not reported | Open |
| 12 | Grok 4.3xAI | 47.3% | Not reported | Closed |
| 13 | Qwen 3.6 Max (preview)Alibaba | 47% | Not reported | Closed |
| 14 | Beam 1.0 501B-A23B NVFP4Reflection AI | 46.9% | Not reported | Open |
| 15 | Nemotron 3 UltraNVIDIA | 44.6% | Not reported | Open |
| 16 | Sage-A1.5InternScience | 44.3% | Not reported | Open |
| 17 | Muse Glimmer 30BMeta | 43.6% | Not reported | Open |
| 18 | Ling 3.0 FlashInclusionAI | 41.2% | Not reported | Open |
| 19 | Hy3 PreviewTencent | 41.2% | Not reported | Open |
| 20 | A.X K2SK Telecom | 41% | Not reported | Open |
| 21 | Ling 3.0 Flash FP8InclusionAI | 40.4% | Not reported | Open |
| 22 | Granite 4.2 30BIBM | 38.8% | Not reported | Open |
| 23 | Mercury 2Inception | 38.4% | Not reported | Closed |
| 24 | Mercury 2.5Inception | 38% | Not reported | Closed |
| 25 | K-EXAONE 2.0LG AI Research | 37.4% | Not reported | Open |
| 26 | Spark-X2.6-4BSparkLLM | 36.8% | Not reported | Open |
| 27 | Granite 4.2 8BIBM | 36.1% | Not reported | Open |
| 28 | LFM2.5-24B-A2BLiquidAI | 32.7% | Not reported | Open |
| 29 | Nemotron 3 Nano Omni 30B A3BNVIDIA | 32% | Not reported | Open |
| 30 | Nemotron 3.5 Lightning 30B A3B NVFP4NVIDIA | 31.4% | Not reported | Open |
| 31 | Agents-A1-4BInternScience | 29.6% | Not reported | Open |
| 32 | LFM3-1.7BLiquidAI | 28.5% | Not reported | Open |
| 33 | Ling 2.6 FlashInclusionAI | 27% | Not reported | Open |
| 34 | MiniCPM5-2BOpenBMB | 26.3% | Not reported | Open |
| 35 | Granite 4.2 3BIBM | 24.1% | Not reported | Open |
According to BenchLM.ai, Sakana Fugu leads the SciCode benchmark with a score of 60.1%, followed by Step 5 Preview (58.9%) and Sakana Fugu-Ultra (58.7%). The top models are clustered within 1.4 points, suggesting this benchmark is nearing saturation for frontier models.
35 models have been evaluated on SciCode. The benchmark falls in the Coding category. BenchAlign v5.8 gives SciCode 10% of the Coding reference weight, so it moves the Coding leaderboard and the overall ranking. Reference weights are relative weights in the calibrated model, not fixed shares of a score.
About SciCode
Year
2024
Tasks
80
Freshness and provenance
Version
SciCode 2024
Refresh cadence
Annual
Staleness state
Refreshing
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does SciCode measure?
SciCode evaluates language models on generating code for realistic scientific research problems across 16 subfields of physics, math, chemistry, biology, and material science. Problems decompose into 338 subproblems requiring domain knowledge recall, scientific reasoning, and precise code synthesis. Based on real scripts from published research.
Which model scores highest on SciCode?
Sakana Fugu by Sakana AI currently leads with a score of 60.1% on SciCode.
How many models are evaluated on SciCode?
35 AI models have published results on SciCode in the BenchLM catalog.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.