DeepSeek V4 Pro (Workers AI)

DeepSeek V4 Pro catches a fair share of bugs (54% [43%–65%]) but makes a poor review seat on cost and noise, and on both counts its cheaper sibling does better.

Start with the noise. It raised 2.9 [2.3–3.4] false alarms per clean PR. Its own sibling, DeepSeek V4 Flash, and GLM-5.3 Flash both sit wholly below that interval, and neither can be told apart from Pro on recall. A reviewer that cries wolf on correct code teaches people to skim its comments, which is how real catches get lost.

Then the bill. The review-call price in the spec sheet assumes a typical reviewer's output, but the measured review came to $0.0643 and took 390s. The difference is output: it wrote several times what that call assumes. A Pro review cost 3.6× a Flash review.

Among closed models, Gemini 4 Argon measurably catches more and its review costs less; the Argon comparison shows where they split. Claude Opus 5.5 and GPT-6.1 Sol are also measurably ahead.

If you want DeepSeek weights in your panel, run V4 Flash instead. DeepSeek's own V4 Pro listing is priced at $1.32 / $3.96 per 1M tokens against $1.32 / $3.96 for this one; it was not in the run. The first open-weight results cover how these models came into the benchmark.