DeepSeek V4 Flash (Workers AI)

Of the two DeepSeek models on Workers AI, V4 Flash is the one to run: 49% [37%–60%] of the known bugs at $0.0178 a review, with recall the benchmark cannot separate from V4 Pro's.

Flash is also the cheaper and quieter of the two. A Pro review cost 3.6× as much, and Flash raised fewer false alarms (1.3 [1.0–1.7] per clean PR, an interval wholly below Pro's). The DeepSeek head-to-head has the bug-by-bug view.

Outside the family it looks weaker. It ranks 5 of 17, and 4 ranked models measurably catch more, GPT-5.6 Sol among them. The cheap closed seats are a closer call. GPT-6 Luna can't be told apart from it on recall, and a Flash review cost 6.3× a Luna one. Its false-alarm interval overlaps Luna's, so noise doesn't decide it.

What it has that Luna doesn't is open weights. Among open models on Workers AI, though, GLM-5.3 Flash is the better cheap pick: also tied on recall, and its review costs less. Choose DeepSeek V4 Flash when you specifically want a DeepSeek model, or want two open-weight makers in one panel.

Some misses are worth knowing: "rpc : cache and reuse compute graphs" went uncaught in every run. The open-weight results post has more on this group of models.