Chinese open AI models rival OpenAI on hacking at a fraction of the cost
What's the deal? Free, downloadable Chinese AI models are closing the gap with OpenAI's closed cyber models on simulated hacking tasks — and costing far less to run, according to research from Vals reported by WIREDDealroom has a profile for this one. Try Dealroom →. In multi-step network hacking tests, DeepSeek scored about 73%, compared with 81% for OpenAI's Daybreak model.
The price gap is the headline. Running those tests cost $2 to $3 with the Chinese model, against $17 to $46 for the OpenAI equivalent. On Vals' separate CyberBench leaderboard, DeepSeek V4.1 Flash reached 73.7% at roughly 11 cents per test.
Why now? Chinese open-weight models are drawing fresh scrutiny for cyber capability. A separate assessment called Z.ai's GLM-5.3 the strongest open-weight cyber model tested, though it still trailed the US frontier on aggregate measures.
What could go wrong? Cheap, downloadable models could broaden access to offensive hacking capability as much as defensive tools. OpenAI has expanded identity-verified access to specialised cyber models, while Anthropic has tightened evaluation safeguards after testing incidents.
Benchmark scores are not proof of real-world results. Simulated performance in a controlled test does not guarantee a model can breach live systems.
The signal: The capability gap between open and closed models is narrowing, and the cost gap is widening in open models' favour. That combination — strong performance at roughly a tenth of the price, with weights anyone can download — reshapes who can access frontier-grade cyber tools.
Read more: WIRED
Image credit: Generated with Gemini
