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
MaxxBench (mutated DeepSWE tasks). Every model the catalog carries a published MaxxBench value for, ranked by that value.
unLAP's MaxxBench takes 89 DeepSWE software engineering tasks and changes the instructions or starting code while keeping the core problem, to test whether coding agents adapt to new requirements or fall back on a memorized solution. 14 models in the ModelScale catalog have a published value.
Each point is a model with a published MaxxBench value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
14 models with a published MaxxBench 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 | Pass rate over 178 agent attempts, mini-swe-agent harness |
|---|---|---|---|
| 1 | Anthropic | 76.4 | |
| 2 | OpenAI | 71.9 | |
| 3 | OpenAI | 69.1 |
What the benchmark measures and how its value is published, from the source’s own record.
MaxxBench (MaxxBench (mutated DeepSWE tasks)) is an external benchmark published by Why MaxxBench (unLAP research post). unLAP's MaxxBench takes 89 DeepSWE software engineering tasks and changes the instructions or starting code while keeping the core problem, to test whether coding agents adapt to new requirements or fall back on a memorized solution. Its evaluation set is described by the source as 89 mutated DeepSWE feature-implementation tasks, two attempts per model. The source rates its difficulty as long-horizon software engineering under changed requirements.
Each model's result is published by Why MaxxBench (unLAP research post) as pass rate over 178 agent attempts, mini-swe-agent harness. ModelScale prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 14 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 | 66.9 |
| 5 | Anthropic | 63.5 |
| 5 | Anthropic | 63.5 |
| 7 | Z.AI | 56.7 |
| 8 | DeepSeek | 55.6 |
| 9 | Z.AI | 55.1 |
| 10 | 53.4 |
| 11 | OpenAI | 51.7 |
| 12 | OpenAI | 46.1 |
| 13 | Meta | 38.8 |
| 14 | DeepSeek | 28.1 |
Evidence key: Observed
Rows are ordered by the value Why MaxxBench (unLAP research post) 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.