DeepSeek V4 Pro Reported to Prematurely Halt in Agentic Coding

karminski3 · x · 2026-08-13

Developer tests reveal that the newly released DeepSeek-V4-Pro-0813 exhibits a notable flaw when using reasoningeffort=max.

In a 50-round long-context Agentic Coding task, the model stopped prematurely around rounds 42-43 in two out of three tests, failing to utilize all available opportunities. Its performance was even outpaced by GLM-5.1. The author speculates the issue might stem from flawed reward mechanisms during the RL phase or context window attention decay.

Original post →

More from coding & agent

coding & agent channel →