DeepSeek-V4-Pro Agent Scores Surge Nearly 50 Points in Internal Tests

DeepSeek-V4-Pro shows massive agent capability gains, with DeepSWE up 49.9 points. It now outperforms Claude Opus 4.8 on multiple benchmarks while maintaining unchanged API pricing.

Woofun AI data shows that DeepSeek-V4-Pro-0813 demonstrates substantial improvements in agent capabilities compared to its preview version. Internal metrics indicate DeepSWE scores jumped from 12.8 to 62.7, a gain of 49.9 points. CyberGym results increased from 52.7 to 83.3, and AutomationBench rose from 12.8 to 31.8.

The updated model exceeds Claude Opus 4.8 in several evaluations, including Terminal Bench 2.1 (87.9 vs 85.0) and CyberGym (83.3 vs 78.3). AutomationBench also surpasses Fable 5 with a score of 31.8 against 29.1. API pricing remains at 3 yuan per million input tokens and 6 yuan for output. These figures derive solely from internal testing, with independent verification pending.