DeepSeek-V4-Pro Official Version Sees Explosive Agent Capability Growth, DeepSWE Jumps Nearly 50 Points in One Go

According to monitoring by Beating, the Agent performance of DeepSeek-V4-Pro-0813 has seen a significant leap. Internal test results leaked by DeepSeek show that DeepSWE skyrocketed from 12.8 in the Preview version to 62.7, an increase of 49.9 points in one go. CyberGym also rose from 52.7 to 83.3, and AutomationBench increased from 12.8 to 31.8. The new version has surpassed Claude Opus 4.8 in multiple evaluations. Terminal Bench 2.1 reached 87.9 compared to 85.0, CyberGym scored 83.3 compared to 78.3, and DeepSWE achieved 62.7 compared to 58.0. AutomationBench even surpassed Fable 5's 29.1 with a score of 31.8. What’s astonishing is that despite the model's upgrade from Preview to 0813, the price hasn’t increased at all. The V4-Pro API still maintains a cost of 3 yuan per million tokens for input and 6 yuan for output. However, these results are currently based on DeepSeek's internal testing, and third-party verification has yet to fully replicate them. Particularly, the nearly 50-point surge in DeepSWE and the Agent evaluation's heavy reliance on Harness mean that the actual improvement will need to be confirmed through external testing. [Original Link]