An open-weight DeepSeek model came close to OpenAI’s restricted cyber model on a new benchmark from evaluation firm Vals AI, at about a tenth of the cost per run, according to Vals’ CyberOffense leaderboard. Vals let AI agents attack two simulated company networks on their own, with no human help.
GPT-6 Sol Daybreak, which Vals says OpenAI offers only through its restricted Daybreak cyber program, finished first. DeepSeek V4.1 Flash 1017, which Vals describes as “an open-weight model that anyone can download,” followed. Vals says DeepSeek scored level with Daybreak on the first network for under $3 per run, against about $22 for Daybreak.
Related Short · Open-Weight AI and Security
An earlier FourWeekMBA Short, “Open-Weight AI Hit 6 Firefox Releases at $1,000 Each.” It does not discuss Vals AI, the CyberOffense benchmark or the models in this piece.
The key insight: Vals AI’s CyberOffense leaderboard puts OpenAI’s restricted GPT-6 Sol Daybreak first and the open-weight DeepSeek V4.1 Flash 1017 close behind, at under $3 per run against about $22, per Vals.
Vals built two simulated networks, which it calls cyber ranges: Nova, with 21 objectives, and Lighthouse, with 31. Each range is isolated from real systems. Vals says it credits an objective only when a hidden marker shows the agent reached that point, and never on the model’s own claim.
Vals compares models at equal compute: a run counts only the objectives completed within 500 million cumulative tokens. The overall score averages the share completed on each range, with both ranges counting equally.
Share of attack objectives each model completed on Vals AI’s two simulated company networks, Nova (21 objectives) and Lighthouse (31). Percentages and overall averages are our arithmetic from the objective counts on Vals’ CyberOffense leaderboard. Vals describes DeepSeek V4.1 Flash 1017 as open-weight and GPT-6 Sol Daybreak as offered only through OpenAI’s restricted Daybreak program.
The Results
On Nova, Vals lists GPT-6 Sol Daybreak at 15.7 of 21 objectives (an average of three runs) and DeepSeek V4.1 Flash 1017 at 16.3 of 21 (also an average of three). On Lighthouse, GPT-6 Sol Daybreak completed 27 of 31, DeepSeek 21 of 31 and Meta Wonder 28 of 31.
Kimi K3.1 Pro completed 13 of 21 and 20 of 31, Grok 4.7 12 and 16, Kimi K3 9 and 10, and GLM 5.3 Flash 5 and 5. Vals says Grok 4.7, Kimi K3 and GLM 5.3 Flash each stalled well before their budgets ran out.
Vals also says GPT-6 Sol Daybreak reached 25 of 31 objectives on the harder range inside 100 million tokens, while DeepSeek “completed nothing in its first 25 million tokens and made most of its progress after 100 million.”
Our Numbers
By our arithmetic, GPT-6 Sol Daybreak’s overall score, the average of its shares on the two ranges, is about 80.9% (74.8% on Nova and 87.1% on Lighthouse), and DeepSeek V4.1 Flash 1017’s is about 72.7% (77.6% and 67.7%), a gap of about 8 points. On Vals’ cost figures, about $22 against under $3 is more than seven times.
Vals frames the result for defenders: “The same models that find a way into a network also find the weaknesses that let them in.”
The Structural Read
As we read the leaderboard, the gap between the two leaders depends on the network: DeepSeek V4.1 Flash 1017 completed more objectives on Nova (16.3 of 21 against 15.7), while GPT-6 Sol Daybreak led on Lighthouse (27 of 31 against 21).
As we read Vals’ methodology, the equal-compute rule matters for cost comparisons: a run counts only objectives completed within 500 million cumulative tokens, and Vals says DeepSeek made most of its progress after 100 million.
As we read the results, the field is uneven: Meta Wonder completed 28 of 31 objectives on Lighthouse but 7 of 21 on Nova, by Vals’ counts.
Vals AI, CyberOffense leaderboard
“The gap that used to separate a restricted frontier model from a freely available one is now small.”
Three Implications
FOR OPENAI Vals lists GPT-6 Sol Daybreak first, a model it says OpenAI offers only through its restricted Daybreak cyber program.
FOR DEEPSEEK Vals describes DeepSeek V4.1 Flash 1017 as an open-weight model that anyone can download, scoring level with Daybreak on Nova for under $3 per run.
FOR DEFENDERS Vals says the same models that find a way into a network also find the weaknesses that let them in.
The framework says: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, Vals’ results bear on which layer is scarce: a restricted model led, and an open-weight one followed closely at lower cost. The framework does not discuss cyber benchmarks.
What Is Not Established
The leaderboard page does not give a publication date; we read it on November 8, 2026. These are results on Vals’ two simulated networks, not real company systems. Vals did not publish per-model results for every run, and the cost comparison is Vals’ own. We have not seen comment from OpenAI or DeepSeek on the results.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
On Vals AI’s CyberOffense test, OpenAI’s restricted GPT-6 Sol Daybreak led, and the open-weight DeepSeek V4.1 Flash 1017 came within about 8 points overall by our arithmetic, at under $3 per run against about $22, per Vals.
A note on sourcing. Results and cost figures are from Vals AI’s CyberOffense leaderboard, read on November 8, 2026. Percentages and overall averages are our arithmetic from Vals’ objective counts. Results are on simulated networks. This piece is not investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
