DeepSeek V4.1 Flash matches restricted GPT-6 Sol Daybreak on one hacking range and trails on the other

DeepSeek's downloadable V4.1 Flash model scored slightly higher than OpenAI's access-gated GPT-6 Sol Daybreak on one of the two simulated corporate networks in Vals AI's CyberOffense benchmark, and lower on the other. That is the whole story, and it is more interesting than either "DeepSeek wins" or "DeepSeek loses." Vals AI, an independent evaluation company, last updated the page on October 23, 2026.

What the benchmark asks an agent to do

CyberOffense gives a model a terminal and asks whether it can compromise a realistic enterprise network on its own. The networks are "cyber ranges," isolated copies of a fake organization. Vals runs two. Nova has 21 objectives on a Windows network with several connected Active Directory domains, and the agent starts from a workstation and a public web application. Lighthouse has 31 objectives across two separate corporate networks, so the agent has to use what it learns in the first to get into the second, through phishing, Active Directory attacks and abuse of domain trusts.

Every model runs in the open-source Pi agent framework on an isolated Kali Linux machine that can reach only the model provider and the range. Each run starts from a fresh copy of the range. Each objective has a hidden marker, and a private answer key checks the agent's command output for it. A model's own claim that it succeeded earns nothing. That design is the best thing about the benchmark: it grades what happened, not what the model said happened.

The scores

Here are the numbers from the Vals page, as of October 23, 2026.

• GPT-6 Sol Daybreak: 15.7 of 21 on Nova (average of three runs, range 12 to 19) and 27 of 31 on Lighthouse.
• DeepSeek V4.1 Flash 1017: 16.3 of 21 on Nova (average of three runs, range 15 to 18) and 21 of 31 on Lighthouse.
• Kimi K3.1 Pro: 13 of 21 and 20 of 31.
• Meta Wonder: 7 of 21 and 28 of 31, the highest single Lighthouse score.
• Grok 4.7: 12 of 21 and 16 of 31.
• Kimi K3: 9 of 21 and 10 of 31.
• GLM 5.3 Flash: 5 of 21 and 5 of 31.

Vals calls GPT-6 Sol Daybreak the overall leader, because the overall score averages the share completed on each range and weights the two equally. DeepSeek's Nova average is higher, but its range of 15 to 18 overlaps Daybreak's 12 to 19 so heavily that with three runs we would not call it a lead. The Lighthouse gap, 21 against 27, is the clearer result.

Cost is the other half

Vals says DeepSeek matched Daybreak on Nova for under $3 per run, versus about $22 for Daybreak. The pattern in the token data differs too. Daybreak reached 25 of 31 Lighthouse objectives within 100 million tokens, while DeepSeek completed nothing in its first 25 million tokens and made most of its progress after 100 million. A slow, cheap model can look worse under a fixed budget than a fast, expensive one, and the page itself warns that the 500 million token cap may understate models that progress slowly but cheaply.

The price list explains the gap. Vals lists DeepSeek V4.1 Flash as an open-weight model released September 10, 2026, at $0.30 per million input tokens and $1.20 per million output tokens. AlphaSignal's October 20 write-up of Vals' preliminary result lists GPT-6 Sol at $2 and $10. On the public CyberBench v1.1 leaderboard, which measures vulnerability discovery and patching, the same write-up reports DeepSeek at 73.69% and $0.07 per test against 77.98% and $1.34 per test for GPT-6 Sol, roughly 19 times the cost for 4.29 points of accuracy.

Why access controls are the real subject

GPT-6 Sol was run through OpenAI's Daybreak trusted-access program, which reduces refusals on approved cybersecurity tasks. DeepSeek's weights can be downloaded and run without anyone's approval. That is the policy point. If a downloadable model gets within a few objectives of a gated one on a task the gated tier exists to control, then a provider-side access program covers less than it appears to. We tracked the wider gap between open and restricted cyber capability in our open-weight cyber capability gap tracker, and our earlier DeepSeek V4.1 Flash 1017 price and benchmark piece covers the model itself.

What the result does not show

Vals lists its own caveats, and they matter. There are only two ranges. Grok 4.7, Kimi K3 and GLM 5.3 Flash stopped on their own after stalling, so their scores show where they ran out of ideas, not a compute limit; GLM spent about 36 hours largely waiting on a background job that never returned. No model completed the final objectives on Lighthouse, and only one run reached the final stage on Nova. The authors say they reviewed every run and found no shortcuts, but the answer key and scoring are not public, so nobody outside Vals can check that claim. AlphaSignal's write-up adds that a simulated range leaves out social engineering, persistence and operational security, and that Vals had not published the scenario, rubric or run logs when that piece ran on October 20.

There is also a version problem. Vals lists "DeepSeek V4.1 Flash" with a September 10 release date and a separate "DeepSeek V4.1 Flash 1017" dated October 17, and its page does not explain how the two relate. The CyberOffense result is for the 1017 build.

Our take

Do not read this as DeepSeek beating OpenAI. Read it as the first independent range test in which a model anyone can download lands within a few objectives of the best access-gated one, at a fraction of the cost. For defenders that is a planning input: assume attackers with a laptop and a modest budget have something close to the gated tier on at least some tasks. For regulators it is a harder question about what gating achieves. We will update this when Vals adds ranges or publishes run logs, and until then the right confidence in the headline is moderate.