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.

How BenchLM shows SciCode right now

BenchLM is tracking SciCode in the local dataset, but exact-source verification records for these rows are still being attached. To avoid a blank benchmark page, BenchLM shows the current tracked rows below as a display-only reference table.

These tracked rows are useful for inspection and spot-checking, but until exact-source attachments are completed they should not be treated as fully verified public benchmark rows.

24 tracked modelsLocal tracked rowsAwaiting exact-source attachmentsDisplay only

Tracked score on SciCode — April 7, 2026

BenchLM mirrors the published tracked score view for SciCode. Gemini 3.1 Pro leads the public snapshot at 59% , followed by Claude Mythos Preview (58.7%) and GPT-5.4 Pro (56.2%). BenchLM does not use these results to rank models overall.

24 modelsCoding10% of category scoreRefreshingUpdated April 7, 2026

The published SciCode snapshot is tightly clustered at the top: Gemini 3.1 Pro sits at 59%, while the third row is only 2.8 points behind. The broader top-10 spread is 13.2 points, so the benchmark still separates strong models even when the leaders cluster.

24 models have been evaluated on SciCode. The benchmark falls in the Coding category. This category carries a 20% weight in BenchLM.ai's overall scoring system. Within that category, SciCode contributes 10% of the category score, so strong performance here directly affects a model's overall ranking.

About SciCode

Year

2024

Tasks

80

BenchLM freshness & provenance

Version

SciCode 2024

Refresh cadence

Annual

Staleness state

Refreshing

Question availability

Public benchmark set

Refreshing

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.

Tracked score table (24 models)

#1
Gemini 3.1 Progemini-3-1-pro
59%
#2
Claude Mythos Previewclaude-mythos-preview
58.7%
#3
GPT-5.4 Progpt-5-4-pro
56.2%
#4
GPT-5.4gpt-5-4
52.5%
#5
Kimi K2.5kimi-k2-5
48.7%
#6
Claude Opus 4.6claude-opus-4-6
48.3%
#7
Gemini 3 Progemini-3-pro
47.2%
#8
GPT-5.3 Codexgpt-5-3-codex
46.2%
#9
GPT-5.2gpt-5-2
45.8%
#10
Qwen3.6 Plusqwen3-6-plus
45.8%
#11
GLM-5 (Reasoning)glm-5-reasoning
44.3%
#12
Claude Sonnet 4.6claude-sonnet-4-6
42.1%
#13
Nemotron 3 Ultra 500Bnemotron-3-ultra-500b
42%
#14
GLM-5glm-5
41.7%
#15
Qwen3.5 397Bqwen3-5-397b
40.2%
#16
GPT-5.3 Instantgpt-5-3-instant
39.1%
#17
Grok 4.1grok-4-1
38.5%
#18
o4-mini (high)o4-mini-high
38.4%
#19
DeepSeek V3.2 (Thinking)deepseek-v3-2-thinking
36.5%
#20
Llama 4 Behemothllama-4-behemoth
35.2%
#21
Mistral Large 3mistral-large-3
32.1%
#22
Step 3.5 Flashstep-3-5-flash
30.8%
#23
MiniMax M2.7minimax-m2-7
29.5%
#24
Llama 4 Maverickllama-4-maverick
28.4%

FAQ

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 leads the published SciCode snapshot?

Gemini 3.1 Pro currently leads the published SciCode snapshot with a tracked score of 59%. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on SciCode?

24 AI models are included in BenchLM's mirrored SciCode snapshot, based on the public leaderboard captured on April 7, 2026.

Last updated: April 7, 2026 · mirrored from the public benchmark leaderboard

Weekly LLM Benchmark Digest

Get notified when new models drop, benchmark scores change, or the leaderboard shifts. One email per week.

Free. No spam. Unsubscribe anytime. We only store derived location metadata for consent routing.