BenchLeader
Menu
Best LLMsChoose a modelLeaderboardsModelsBenchmarksProvidersReleases12

FrontierCode

Hard real-world coding tasks run in agent harnesses (Cognition).

As of 24 Oct 2026, Claude Opus 5.5 leads FrontierCode on BenchLeader with 54.6%, ahead of Claude Fable 5 at 53.5%, across 45 model configurations with a published result.

Published by
Cognitiondata via Epoch AI Benchmarking Hub
Category
Coding
Index weight
0.5
Models
45
Data as of
24 Oct 2026
Newest result
7 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

FrontierCode gives agents hard real-world programming tasks, closer to a week's ticket than a bug fix, inside an agent harness with a repository and tools, and checks the result with tests. Maintained by Cognition.

How it is scored

Percent of tasks passing their tests in the published harness.

What to keep in mind

Harness-specific and maintained by a company that also sells an agent, which the publisher discloses. Task count is modest, so treat gaps of a few points as noise.

Top 15
  1. 1Claude Opus 5.5 (medium)54.6%
  2. 2Claude Fable 553.5%
  3. 3Claude Opus 5 (max)53.4%
  4. 4GPT-6 Astra (max)53.3%
  5. 5Claude Sonnet 5.5 (xhigh)52.1%
  6. 6Claude Fable 5.1 (medium)50.9%
  7. 7GPT-6.1 Sol (medium)50.2%
  8. 8SWE-2 (max)50.0%
  9. 9GPT-6 Sol (max)49.3%
  10. 10Grok 4.648.0%
  11. 11Grok 4.747.6%
  12. 12Gemini 4 Argon (medium)47.5%
  13. 13GPT-5.6 Sol47.5%
  14. 14Claude Opus 4.846.5%
  15. 15Claude Haiku 5.5 (max)46.4%
45 of 45
#
1Claude Opus 5.5medium54.6%66.52026-09-22
2Claude Fable 553.5%67.12026-06-09
3Claude Opus 5max53.4%68.82026-07-24
4GPT-6 Astramax53.3%70.42026-09-03
5Claude Sonnet 5.5xhigh52.1%67.92026-09-28
6Claude Fable 5.1medium50.9%67.32026-09-01
7GPT-6.1 Solmedium50.2%67.22026-09-29
8SWE-2max50.0%–2026-09-10
9GPT-6 Solmax49.3%66.42026-09-22
10Grok 4.648.0%57.82026-08-12
11Grok 4.747.6%–2026-09-21
12Gemini 4 Argonmedium47.5%–2026-09-30
13GPT-5.6 Sol47.5%55.92026-07-09
14Claude Opus 4.846.5%60.12026-05-28
15Claude Haiku 5.5max46.4%60.52026-10-07
16Kimi K344.2%59.12026-07-16
17DeepSeek V4.1 Flashhigh43.9%55.62026-09-09
18Gemini 3.7 Flash43.6%53.42026-08-13
19GPT-5.543.0%60.32026-04-23
20Claude Sonnet 542.7%56.12026-06-30
21Grok 4.542.4%59.02026-07-08
22GPT-6 Lunamax42.4%60.02026-09-22
23SWE-1.742.0%–2026-07-08
24GPT-5.6 Terra41.3%52.72026-07-09
25Gemini 3.8 Flashmedium41.2%64.42026-09-02
26GLM 5.3max40.1%63.52026-08-14
27GPT-5.6 Luna39.8%48.02026-07-09
28Mistral Large 4high39.4%49.82026-10-06
29Claude Opus 4.738.5%63.32026-04-16
30Gemini 3.6 Flash34.4%49.52026-07-21
31GLM 5.3 Flashmax31.8%54.22026-08-20
32Kimi K2.7 Code30.1%55.82026-06-12
33DeepSeek V4 Prohigh28.6%57.52026-08-13
34GPT-5.4 Mini27.0%–2026-03-17
35Claude Opus 4.626.6%62.62026-02-05
36Composer 2.525.6%–2026-05-18
37GLM 5.2no reasoning24.5%48.02026-06-16
38Claude Sonnet 4.624.3%57.12026-02-17
39DeepSeek V4 Flash18.8%52.22026-07-31
40DeepSeek V4 Prono reasoning17.6%46.92026-04-24
41MiniMax M314.7%56.42026-06-01
42Nemotron 3 Ultra13.6%44.12026-06-04
43Qwen3.7 Plus10.2%58.52026-06-02
44SWE-1.69.4%–2026-04-07
45Mistral Medium 3.58.0%49.42026-04-28

More coding benchmarks · All LLM benchmarks

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

FrontierCode: questions

What does FrontierCode measure?
FrontierCode gives agents hard real-world programming tasks, closer to a week's ticket than a bug fix, inside an agent harness with a repository and tools, and checks the result with tests. Maintained by Cognition. Scores are reported in percent of tasks solved; higher is better.
Which AI model leads FrontierCode?
Claude Opus 5.5 leads FrontierCode with 54.6% as of 24 Oct 2026, ahead of Claude Fable 5 at 53.5%.
How many models have FrontierCode results?
45 model configurations have a FrontierCode result on BenchLeader, all taken from Cognition via Epoch AI Benchmarking Hub.
Who runs FrontierCode and how often is it updated?
FrontierCode is published by Cognition. BenchLeader re-reads the published results once a day and records the date each result was published.
Does FrontierCode count toward the BenchLeader Index?
Yes. FrontierCode contributes to the coding category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.