A free DeepSeek model nearly matched OpenAI’s gated cyber model at a tenth of the cost

An open-weight model that anyone can download now comes close to OpenAI’s restricted cyber model at breaking into simulated company networks. Wired reported on the results from the evaluation firm Vals: DeepSeek completed 73% of objectives on average, against 81% for OpenAI’s Daybreak.

Two simulated ranges, seven models

Vals published the results on its CyberOffense benchmark page, last updated on October 23. The test runs an open-source agent loop on an isolated Kali Linux machine that can reach only the model provider and the target network. Scoring relies on hidden markers checked against a private answer key in the recorded command output, so a model’s own claims earn nothing.

The first range, called Nova, links several Active Directory domains and starts from a workstation and a public web application. It has 21 objectives. The second, Lighthouse, has 31 objectives and consists of two separate corporate networks, where information taken from the first is the only way into the second.

On Nova, DeepSeek V4.1 Flash 1017 averaged 16.3 objectives over three runs, against 15.7 for GPT-6 Sol Daybreak, which Vals calls level. On Lighthouse, Daybreak completed 27 and DeepSeek 21. Meta’s Wonder model scored 28 on Lighthouse but only 7 on Nova. No model finished the final Lighthouse objectives.

Each run was capped at 500 million tokens or 48 hours, and Vals says it reviewed every run and found no shortcuts. As the Vals authors put it, “Cyber capability that was limited to trusted-access AI is now available to anyone.”

Under $3 a run against about $22

Cost is where the gap is widest. Vals puts DeepSeek at under $3 per run and Daybreak at about $22, roughly a tenth. Wired quotes a wider range of $17 to $46 for the OpenAI model, against $2 to $3 for DeepSeek. Open weights can run on a company’s own hardware, which Vals CEO Rayan Krishnan told Wired gives defenders a strong reason to prefer them.

DeepSeek was slower and less efficient. Vals reports it completed nothing in its first 25 million tokens and needed two to four times the tokens Daybreak used to reach a similar point. Daybreak had 25 of the 31 Lighthouse objectives within 100 million tokens and finished in under three hours.

How this fits earlier measurements

The result lands after the UK AI Security Institute’s July assessment, as summarized by Winbuzzer. AISI found open-weight models trailing closed ones by about four to seven months, down from six to ten months through much of 2025. In its tests, a 100-million-token run cost about $85 on Claude Opus 4.6, $46 on GLM-5.2 and $1.19 on DeepSeek V4-Pro.

AISI’s setup differed, with no live attacks or active defenders. Vals adds a head-to-head against a restricted model, and the gap there is a few points rather than months.

Not every open model did well. On the Vals table, GLM 5.3 Flash completed 5 of 21 and 5 of 31 objectives, and Vals notes Grok 4.7, Kimi K3 and GLM stalled and ended their own runs. Kimi K3.1 Pro, another open-weight Chinese model, completed 13 of 21 and 20 of 31.

What the tests do not settle

Vals lists its own limits: only two ranges, a specific agent setup, and a token cap that may understate slow, cheap models. The run counts are also uneven, with some Nova scores averaged over three runs and others from a single run.

Wired also cites security experts who caution that attackers have mostly used older, weaker models so far. Alex Zenla of Edera told the outlet that closed labs likely hold unreleased models beyond these results, and called it a cat-and-mouse game. OpenAI did not immediately respond to Wired’s request for comment.

For defenders the point is practical: strong offensive-security tooling no longer depends on trusted-access programs. Tools that strip safety training from open models, like those in our report on Heretic 2.0, are part of why Zenla flags abliterated models as a concern. Other open releases are also being judged on cyber ability, as in our look at Mistral’s Le Chonk weights.