Get started

The Kotlin Benchmark

The Kotlin Benchmark is built to evaluate models on realistic, idiomatic Kotlin engineering tasks across multiple dimensions.

Setup (Agent + LLM) proprietarySubmittedResolvedResolution rate (%)Avg. tokens (M)Avg. latencyDate
1Claude Code + Opus 4.7 xhigh
1059085.7110.5910h 37m06.05.2026
2Junie (Opus 4.7 max)
1058681.9019.9819h 58m23.04.2026
3Codex + GPT 5.5 xHigh
1058681.9038.788h 30m26.04.2026
4Claude Code + Opus 4.6 max
1058480.0013.9312h 39m19.03.2026
5Codex + GPT 5.3 Codex xHigh
1058278.108.568h 1m19.03.2026
6Junie (Opus-4.6)
1058177.147.458h 36m23.03.2026
7Codex + GPT 5.4 xHigh
1058177.1410.859h 33m19.03.2026
8Claude Code + Opus 4.7 medium
1058076.195.314h 54m07.05.2026
9Codex + GPT 5.4 medium
1058076.198.296h 25m19.03.2026
10Claude Code + Sonnet 4.6 high
1057874.2912.1713h 49m21.04.2026
11Claude Code + Opus 4.6 medium
1057571.438.407h 43m18.03.2026
12Codex + GPT 5.3 Codex medium
1057268.576.325h 26m19.03.2026
13Junie (GPT-5.4)
1057268.5715.2911h 30m02.04.2026
14Claude Code + Sonnet 4.6 medium
1057066.6710.1910h 28m19.03.2026
15Junie (GPT-5.3-codex)
1056965.7115.6811h 8m23.03.2026
16Gemini CLI (Gemini 3.1 Pro)
1056965.7121.0317h 1m21.03.2026
17Junie (Gemini 3 Flash)
1056460.9540.1711h 27m23.03.2026
18Junie (Gemini 3.1 Pro Preview)
1056360.0032.4113h 41m23.03.2026
19Gemini CLI (Gemini 3 Flash)
1054744.7636.5311h 28m23.03.2026

How Kotlin Benchmark works

See how the benchmark is built — the dataset of real Kotlin tasks, the harness that runs them, and how every result is scored.