Fireworks AI says DeepSeek V4 Pro 0813 completed security tasks in its CyberGym evaluation, which uses real vulnerable codebases and checks whether an exploit works and a patch is made. Across a 697-task comparison, Fireworks reports that the model solved more tasks than GPT-5.5 and cost one-quarter as much per solved vulnerability.
Against Kimi K3, DeepSeek V4 Pro had a lower reported cost per success, but K3 solved more tasks overall. Fireworks also says DeepSeek V4 Pro made tool calls on all 840 traced tasks without refusals or output truncation. These are company-run results; the source provides no independent validation.
Why it matters
For security teams weighing automated code audits, Fireworks’ results suggest DeepSeek V4 Pro may cost less per solved vulnerability than GPT-5.5, but the company’s evaluation has not been independently validated.
