ProgramBenchBrowse 296

ProgramBench

Can the model complete substantial programming work under this benchmark's agent setup?

Version not specifiedExact reported variant
Coding12 ranked models17 reported values4 reportsHigher is better

Top rankings

One row per model, using its best reported score across effort settings.

12
RankModelBest scoreBest reported setting
1Kimi K3Moonshot AI77.8%max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source
2GPT-5.6 SolOpenAI77.6%max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source
3Claude Fable 5Anthropic76.8%max effortMoonshot AI report1 value · 1 reportJul 17, 2026 · source
4Claude Opus 4.8Anthropic71.9%max effortMoonshot AI report3 values · 3 reportsJul 17, 2026 · source
5GPT-5.5OpenAI70.8%xhigh effortMoonshot AI report3 values · 3 reportsJul 17, 2026 · source
6GLM-5.2Z.ai63.7%max effortMoonshot AI report2 values · 2 reportsJul 17, 2026 · source
7Kimi K2.7 CodeMoonshot AI53.6%thinking effortMoonshot AI report1 value · 1 reportJul 22, 2026 · source
8GLM-5.1Z.ai50.9%Reported configurationZ.ai report1 value · 1 reportJun 16, 2026 · source
9Kimi K2.6Moonshot AI48.3%thinking effortMoonshot AI report1 value · 1 reportJul 22, 2026 · source
10DeepSeek V4 ProDeepSeek47.8%Reported configurationZ.ai report1 value · 1 reportJun 16, 2026 · source
11Gemini 3.1 ProGoogle39.5%Reported configurationZ.ai report1 value · 1 reportJun 16, 2026 · source
12DeepSeek V4.1 FlashDeepSeek20.3%Reported configurationDeepSeek report1 value · 1 reportSep 10, 2026 · source

Effort curve

Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.

0
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology

Can the model complete substantial programming work under this benchmark's agent setup?

Software implementation, debugging, or repository work under the published evaluation protocol. Higher is better. The value is the percentage reported in this lab's table.

Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.

DeepSeek's official V4.1-Flash release table; exact source benchmark label ProgramBench retained. Kimi Code harness. Claude Code 2.1.156; max reasoning effort for GLM-5.2. Kimi Code CLI with thinking enabled, temperature 1.0, top-p 0.95, and 262,144-token context; GPT-5.5 used Codex xhigh and Claude Opus 4.8 used Claude Code xhigh.

Method / source