Benchmarks › SciCode

SciCode

By SciCode team (UIUC, Argonne National Laboratory, Carnegie Mellon and others) ↗Results by Artificial Analysis ↗Since 2024Paper ↗
Write Python code to solve scientific computing problems
66.9top score · Claude Opus 5.5
140models ranked
86open-weights models
Sub-problem pass@1higher is better
Oct 20, 2026last updated

How the test works

1

Task

288 subproblems

  • What the model doesWrite Python code to solve scientific computing problems
  • Kinds of tasksScientific programming
  • In → outText → Python code
2

Run

3 runs per task

  • Token budgetNon-reasoning: up to 16,384 output tokens, adjusted to model limits; reasoning: model-specific maximum
  • Temperature0 for non-reasoning; 0.6 for reasoning, unless the model lab recommends otherwise
3

Score

Sub-problem pass@1 (solved on the first try)

  • Graded byPython code execution / unit tests
  • Better isHigher
Sources: artificialanalysis.ai

SciCode leaderboardtop 10 of 140

Full leaderboard with filters →
ModelSciCodePer taskPriceSpeedSupportsContext
Claude Opus 5.5Anthropic
66.9max
$5.98$4 / $20961M
Claude Fable 5.1Anthropic
63.1max
$7.63$10 / $50691M
Muse Wonder FullMeta
62.6
——1M
Gemini 4 ArgonGoogle
61.8high
$1.99$2 / $10841M
Claude Fable 5Anthropic
61max
$8.75$10 / $50671M
Claude Sonnet 5.5Anthropic
61max
$5.46$2 / $101371M
MiMo v2.6 ProXiaomiOpen
60.9thinking
$0.13$0.43 / $0.87541M
Muse WonderMeta
60.3max
$4.52$4 / $201611M
Muse Spark 1.3Meta
59.7xhigh
$1.37$1.25 / $4.252731M
Kimi K3Moonshot AIOpen
59.5max
$2.00$3 / $151121M

SciCode questions

What is SciCode?›

Write Python code to solve scientific computing problems. It has 288 subproblems. Created by SciCode team (UIUC, Argonne National Laboratory, Carnegie Mellon and others).

Which model is best on SciCode?›

Claude Opus 5.5 leads SciCode with 66.9, ahead of Claude Fable 5.1 (63.1) and Muse Wonder Full (62.6).

What is the best open-weights model on SciCode?›

MiMo v2.6 Pro with 60.9, then Kimi K3 (59.5).

How is SciCode scored?›

Sub-problem pass@1 (solved on the first try); python code execution / unit tests; 3 runs per task per task.