DeepSeek V4 Pro 0813 now available via API

DeepSeek has made V4 Pro 0813 available via API on OpenRouter, the latest iteration in its Pro model line. However, the release lacks a formal announcement from DeepSeek itself—the author had to link to OpenRouter because "DeepSeek don't have any obvious announcement page for their new model." This follows a pattern established by earlier releases: DeepSeek published open weights for both the April V4-Pro and the July V4-Flash-0731 model, suggesting the company may eventually open-source V4 Pro 0813 as well, though no confirmation has been made. One technical distinction the author noted is that the model produces visibly different outputs at different reasoning levels. When testing with low, medium, and high reasoning settings, the author received "very different looking pelicans" for each level—a behavior the author has not seen in other models, suggesting a potentially novel approach to how reasoning level affects output generation. Benchmark data for the model has not been officially published by DeepSeek but has instead circulated informally through community channels: first shared in the Official DeepSeek WeChat Group, then posted to Reddit (where moderators deleted it for being "low-effort"), and subsequently converted into an ASCII-art table on Hacker News.