DeepSeek-V4-Pro-0813 posted sharply higher agent benchmark results in a leaked DeepSeek self-test table, with the biggest move in DeepSWE, which climbed to 62.7 from 12.8. CyberGym rose to 83.3 from 52.7 and AutomationBench increased to 31.8 from 12.8. The new version was also shown outperforming Claude Opus 4.8 on several tests, including Terminal Bench 2.1 at 87.9 versus 85.0, CyberGym at 83.3 versus 78.3, and DeepSWE at 62.7 versus 58.0. AutomationBench reached 31.8, above Fable 5 at 29.1. Despite the apparent performance gains from the Preview version to 0813, DeepSeek kept V4-Pro API pricing unchanged at 3 yuan per million input tokens and 6 yuan per million output tokens. The results remain internal testing, however, and full third-party replication has not yet been completed. That caveat may matter especially for DeepSWE, where the nearly 50-point jump is unusually large and agent evaluations can depend heavily on the harness (test-running framework) used.