Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
external benchmark
CompileBench (building real projects from source, pass@1). Every model the catalog carries a published CompileBench value for, ranked by that value.
Quesma's CompileBench hands an agent the unmodified source of an open-source project such as curl, jq or GNU coreutils and asks for a working binary. Tasks include static and musl builds, cross-compiling to Windows and ARM64, and reviving code from 2003. 17 models in the ModelScale catalog have a published value.
Each point is a model with a published CompileBench value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
17 models with a published CompileBench value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Agent in a Docker Linux terminal builds a project from source; the binary is checked |
|---|---|---|---|
| 1 | Anthropic | 86.7 | |
| 1 | OpenAI | 86.7 | |
| 1 | OpenAI | 86.7 |
What the benchmark measures and how its value is published, from the source’s own record.
CompileBench (CompileBench (building real projects from source, pass@1)) is an external benchmark published by CompileBench: can AI compile 22-year-old code? (Quesma). Quesma's CompileBench hands an agent the unmodified source of an open-source project such as curl, jq or GNU coreutils and asks for a working binary. Tasks include static and musl builds, cross-compiling to Windows and ARM64, and reviving code from 2003. Its evaluation set is described by the source as 15 build tasks on cowsay, jq, GNU coreutils and curl, three attempts each (45 attempts). The source rates its difficulty as dependency resolution, legacy toolchains, static linking and cross-compilation.
Each model's result is published by CompileBench: can AI compile 22-year-old code? (Quesma) as agent in a Docker Linux terminal builds a project from source; the binary is checked. ModelScale prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 17 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.
| 4 | Anthropic | 84.4 |
| 5 | Anthropic | 73.3 |
| 6 | xAI | 71.1 |
| 7 | Anthropic | 68.9 |
| 8 | DeepSeek | 64.4 |
| 8 | xAI | 64.4 |
| 10 | xAI | 55.6 |
| 11 | 53.3 |
| 11 | OpenAI | 53.3 |
| 11 | Alibaba | 53.3 |
| 14 | Z.AI | 48.9 |
| 15 | OpenAI | 46.7 |
| 16 | OpenAI | 42.2 |
| 17 | 40.0 |
Evidence key: ObservedLast good
Rows are ordered by the value CompileBench: can AI compile 22-year-old code? (Quesma) published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.