BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

SciCode

Research-grade scientific coding problems.

As of 24 Oct 2026, Claude Opus 5.5 leads SciCode on BenchLeader with 66.9%, ahead of Claude Opus 5.5 at 65.0%, across 216 model configurations with a published result.

Published by
SciCodedata via Epoch AI Benchmarking Hub
Category
Coding
Index weight
0.5
Models
216
Data as of
24 Oct 2026
Newest result
15 Oct 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

SciCode gives models research-grade programming problems from physics, chemistry, biology and materials science, each broken into sub-steps that build a working scientific simulation or analysis.

How it is scored

Percent of main problems fully solved, meaning every sub-step passes its tests. Epoch AI runs the models with the standard prompt.

What to keep in mind

The all-or-nothing scoring makes it harsh: partial credit on sub-steps is not shown. It rewards scientific background knowledge as much as coding.

Top 15
  1. 1Claude Opus 5.5 (max)66.9%
  2. 2Claude Opus 5.5 (xhigh)65.0%
  3. 3Claude Fable 5.1 (max)63.1%
  4. 4Gemini 4 Argon (high)61.8%
  5. 5Claude Fable 5 (max)61.0%
  6. 6Claude Sonnet 5.5 (max)61.0%
  7. 7Claude Fable 5.1 (xhigh)60.9%
  8. 8MiMo-V2.6-Pro60.9%
  9. 9Claude Opus 5.5 (high)60.4%
  10. 10Muse Wonder (max)60.3%
  11. 11Gemini 3.7 Flash (medium)59.8%
  12. 12Muse Spark 1.3 (xhigh)59.7%
  13. 13Kimi K3 (max)59.5%
  14. 14Claude Opus 5.5 (medium)59.3%
  15. 15GLM 5.3 (max)59.0%
216 of 216
#
1Claude Opus 5.5max66.9%72.62026-09-22
2Claude Opus 5.5xhigh65.0%70.02026-09-22
3Claude Fable 5.1max63.1%70.52026-09-01
4Gemini 4 Argonhigh61.8%70.82026-09-30
5Claude Fable 5max61.0%69.32026-06-09
6Claude Sonnet 5.5max61.0%69.32026-09-28
7Claude Fable 5.1xhigh60.9%69.52026-09-01
8MiMo-V2.6-Pro60.9%64.02026-09-21
9Claude Opus 5.5high60.4%69.42026-09-22
10Muse Wondermax60.3%67.52026-10-12
11Gemini 3.7 Flashmedium59.8%63.12026-08-13
12Muse Spark 1.3xhigh59.7%64.92026-09-02
13Kimi K3max59.5%64.72026-07-16
14Claude Opus 5.5medium59.3%66.52026-09-22
15GLM 5.3max59.0%63.52026-08-14
16Step 5 Preview58.9%60.42026-09-18
17Muse Spark 1.3max58.8%67.12026-09-02
18Muse Spark 1.158.8%62.52026-07-09
19Muse Spark 1.1xhigh58.8%61.62026-07-09
20Gemini 3.1 Pro58.7%60.62026-02-19
21Kimi K358.7%59.12026-07-16
22Claude Opus 5.5low58.6%64.72026-09-22
23NewMuse Spark Open 816B A42Bxhigh58.5%58.62026-10-15
24GPT-5.6 Solhigh57.8%66.32026-07-09
25Grok 4.7high57.8%63.92026-09-21
26Claude Fable 5.1high57.6%69.42026-09-01
27GPT-6 Solmax57.6%66.42026-09-22
28GPT-5.6 Solmedium57.4%64.42026-07-09
29Muse Spark 1.2xhigh57.4%63.22026-08-05
30Grok 4.7xhigh57.4%61.62026-09-21
31Claude Sonnet 5.5xhigh57.3%67.92026-09-28
32Gemini 3.7 Flashhigh57.2%63.52026-08-13
33GPT-5.6 Solmax57.1%68.02026-07-09
34GPT-5.6 Solxhigh57.1%66.62026-07-09
35GPT-5.4xhigh56.6%64.72026-03-05
36Gemini 3.8 Flashhigh56.6%64.42026-09-02
37GPT-6 Astramax56.5%70.42026-09-03
38Grok 4.6high56.5%62.92026-08-12
39Claude Opus 5max56.4%68.82026-07-24
40GPT-5.6 Sollow56.4%62.22026-07-09
41GPT-5.5high56.1%65.32026-04-23
42Grok 4.6medium55.9%64.22026-08-12
43GPT-6.1 Solhigh55.8%67.72026-09-29
44GPT-5.5xhigh55.8%66.12026-04-23
45GPT-6 Astraxhigh55.7%68.92026-09-03
46GPT-6.1 Solxhigh55.7%67.72026-09-29
47Claude Fable 5.1low55.7%65.02026-09-01
48GPT-6 Astrahigh55.4%69.62026-09-03
49Claude Fable 5.1medium55.3%67.32026-09-01
50Gemini 3.8 Flashmedium55.1%64.42026-09-02
51GPT-6 Solxhigh55.1%64.32026-09-22
52Claude Opus 5xhigh55.0%67.42026-07-24
53GPT-5.6 Terramax55.0%64.42026-07-09
54Gemini 3.8 Flashlow55.0%61.02026-09-02
55Claude Haiku 5.5max55.0%60.52026-10-07
56Grok 4.5high55.0%59.32026-07-08
57GPT-6 Solhigh54.9%63.72026-09-22
58Grok 4.7low54.9%61.02026-09-21
59GPT-6 Lunamax54.6%60.02026-09-22
60GPT-5.5medium54.5%62.92026-04-23

More coding benchmarks · All LLM benchmarks

Cite as: BenchLeader, “SciCode leaderboard”, data as of 24 Oct 2026, https://www.benchleader.com/benchmarks/scicode.

SciCode: questions

What does SciCode measure?
SciCode gives models research-grade programming problems from physics, chemistry, biology and materials science, each broken into sub-steps that build a working scientific simulation or analysis. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads SciCode?
Claude Opus 5.5 leads SciCode with 66.9% as of 24 Oct 2026, ahead of Claude Opus 5.5 at 65.0%.
How many models have SciCode results?
216 model configurations have a SciCode result on BenchLeader, all taken from SciCode via Epoch AI Benchmarking Hub.
Who runs SciCode and how often is it updated?
SciCode is published by SciCode. BenchLeader re-reads the published results once a day and records the date each result was published.
Does SciCode count toward the BenchLeader Index?
Yes. SciCode contributes to the coding category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.