AI data firm Datacurve has added Kimi K3 to DeepSWE, a benchmark for long-horizon software engineering. The test measures whether models can independently complete real, complex, multi-step development tasks.
Kimi K3 achieved a task pass rate of 69%. GPT-5.6 Sol scored 73%, while Claude Fable 5 managed 70%.
Datacurve says Kimi K3 is the first open-weight model to reach the level of frontier closed-source models on DeepSWE.
DeepSWE includes 113 original tasks drawn from 91 active open-source projects across five programming languages. All models use the same tool set. Tasks are not directly adapted from existing code commits, reducing the risk of training data contamination.
Kimi K3 costs an average of $4.65 per task. GPT-5.6 Sol costs $8.39, and Claude Fable 5 costs $21.63.
https://twitter.com/datacurve/status/2078189882707730535?s=46