A freely downloadable AI model can now do almost everything OpenAI's restricted cyber model can do inside a simulated corporate network — for about a tenth of the cost. In new tests from benchmarking firm Vals, published on November 6 and reported by WIRED and The Decoder, OpenAI's GPT-6 Sol completed 80.8 percent of attack objectives across two mock enterprise networks. DeepSeek's open-weight V4.1 Flash finished at 72.8 percent.

The gap matters because of who can use each model. GPT-6 Sol is available only through OpenAI's Daybreak cyber programme, which is offered to vetted partners. DeepSeek V4.1 Flash can be downloaded by anyone and run on their own hardware. "Cyber capability that was limited to trusted-access AI is now available to anyone," Vals wrote in its results.

How the test worked

Vals' CyberOffense benchmark sets each model loose as an autonomous agent. It runs inside the open-source Pi agent on an isolated Kali Linux machine, starts with nothing but a public web address, and gets no human help at all. Each run has a 48-hour limit and a 500-million-token budget, and an objective counts only when hidden markers appear in the command output — never on the model's own claim.

The benchmark uses two simulated ranges. Nova is a 21-objective Windows network spanning several Active Directory domains. Lighthouse is harder: 31 objectives across two companies, where the agent must take over the first network and use what it finds there to break into the second.

DeepSeek led on one network, trailed on the harder one

On Nova, DeepSeek actually edged ahead, averaging 16.3 of 21 objectives against Daybreak's 15.7 — for under $3 a run, against about $22 for the OpenAI system. On Lighthouse, Daybreak pulled away, completing 27 of 31 objectives in under three hours: it took administrator control of the first company, stole passwords, and used them to break into the second. DeepSeek managed 21.

The other models finished further back. Meta's Wonder completed 28 Lighthouse objectives, the most on that range, while Kimi K3.1 Pro reached 20, Grok 4.7 reached 16, Kimi K3 reached 10, and Z.ai's GLM 5.3 Flash stalled at five objectives on each range. No model completed either full attack chain.

DeepSeek's score came with a quirk: it scored nothing in its first 25 million tokens and made most of its progress only after passing 100 million. Vals concluded that short, compute-capped tests understate cheap open models, which keep improving the longer they are allowed to run.

The gap is closing fast

The findings line up with an analysis published in July by the UK's AI Security Institute, which put the gap between open and closed frontier models on cyber capability at four to seven months, down from six to ten months for most of 2025. The institute estimated a cyber-range run at $1.19 with DeepSeek V4-Pro, against $85 with Anthropic's Opus 4.5.

The timing is awkward for the labs. OpenAI released GPT-6 Cyber on November 5, rating it "Critical" for cybersecurity and limiting it to approved Daybreak Red customers. But open weights cannot be gated once they are released. Vals researcher Nikil Ravi told WIRED that the best models locked inside the labs are "starting to be matched by open models that anybody can download."

Vals expects attacks on real networks to increase in the coming months. The same results cut both ways: the models that find and exploit vulnerabilities can also help companies find and fix them first — if defenders move before attackers do.

By Hamza Faisal