DeepSeek V4.1 Flash vs GPT-6 Sol Daybreak: A Free Model Is 8 Points Behind OpenAI's Restricted Cyber AI

In Vals AI's CyberOffense benchmark, OpenAI's restricted GPT-6 Sol Daybreak completed about 81 percent of attack objectives across two simulated enterprise networks, while the freely downloadable DeepSeek V4.1 Flash completed about 73 percent. The gap is eight points, and the cost gap is far wider: under 3 dollars a run for DeepSeek against about 22 dollars for Daybreak. Those two numbers, published by Vals on 6 November 2026 and reported by WIRED, are the clearest sign yet that the cyber capability gap between gated frontier models and open-weight models is closing fast.

Before going further, read the warning that belongs at the top of any honest coverage of this story. These were simulated networks, not real companies.

What Vals actually tested

A cyber range is a simulated network built to look like a real organisation, with servers, workstations, user accounts and realistic security gaps, isolated from real systems so an agent can attack it safely. Vals built two of them.

Each model ran as an autonomous agent in the open-source Pi agent on an isolated Kali Linux machine, a standard environment for penetration testing. The agent started with nothing but the address of a public web application and the authorised scope, then worked alone until it finished, hit its token budget, or reached the 48-hour limit.

Seven models took the test.

Illustration: Analyst seen from behind at a desk with four monitors showing network topology maps in a dark security operations centre

The scoreboard: where each model won and lost

The overall score averages the share of objectives completed on each range, with both ranges counting equally. That weighting matters, because the two networks tell different stories. On Nova, DeepSeek actually edged ahead, averaging 16.3 of 21 objectives across three runs against 15.7 for Daybreak. On Lighthouse, Daybreak pulled away, completing 27 of 31 objectives in under three hours, against 21 for DeepSeek.

Model	Nova (of 21)	Lighthouse (of 31)	Overall	Cost per run
GPT-6 Sol Daybreak	15.7 average	27	80.8 percent	About 22 dollars
DeepSeek V4.1 Flash 1017	16.3 average	21	72.8 percent	Under 3 dollars
Meta Wonder	Not reported	28	Not reported	Not reported
Kimi K3.1 Pro	Not reported	20	Not reported	Not reported
Grok 4.7	Not reported	16	Not reported	Not reported
GLM 5.3 Flash	5	5	Not reported	Not reported

Two details in that table deserve attention. First, Meta Wonder completed the most Lighthouse objectives of any model, 28 of 31, ahead of Daybreak's 27.

The models also tied on some tasks, as Vals and WIRED both note, and the two leaders were level on the first network. That is why the headline gap of eight points understates how close the contest was on Nova and overstates it as a single number.

The cost gap is the real story

Capability was close.

Vals chief executive Rayan Krishnan told WIRED that the price difference gives defenders a strong incentive to use open models rather than proprietary ones. The logic is straightforward: the same capability that finds a way into a network also finds the weaknesses that let it in, and a 3-dollar run is a test almost any security team can afford to repeat.

That framing is exactly why this benchmark matters beyond the leaderboard. OpenAI, Anthropic and other labs have justified restricted releases on the argument that gating buys defenders time before advanced hacking capability becomes broadly available.

Why DeepSeek was slow to start

DeepSeek started slowly: it scored nothing in its first 25 million tokens, and Vals found it made most of its progress only after 100 million tokens. That slow start is the inefficiency the rest of this section is about.

Illustration: Network engineer with a laptop walking between server racks in a data center corridor

That inefficiency barely mattered to the final bill, because DeepSeek's tokens are so cheap that running it for a long time still cost under 3 dollars. But it matters a great deal to how other benchmarks should be read. Vals concluded that short, compute-capped tests understate cheap open models. An evaluation that capped compute early would have ranked DeepSeek far below its real reach. Any benchmark that gives every model the same small token budget will flatter efficient, expensive models and punish cheap ones that need more room to work. Future tests that want a fair comparison will need to report results at several compute levels, not one.

There is a second lesson in the token numbers. More compute did not help the weaker models. Grok 4.7, Kimi K3 and GLM 5.3 Flash each stalled well before their budgets ran out and never recovered, and every run in the set ended on its own rather than at the token or time limit. Each score therefore marks where a model ran out of ideas, not where Vals stopped it. That is a cleaner result than a test cut short by the clock, and it makes the gap between DeepSeek and the models behind it more meaningful.

Where GLM and the other models landed

Z.ai's GLM 5.3 Flash came last of the seven models, completing five objectives on each range. Vals called it the clearest dead end: on one range it spent about 36 hours and a thousand turns, much of it sleeping and polling for a background job that never returned, and still finished at five objectives.

Those are different tests with different harnesses, and their scores cannot be compared with Vals' results. The honest reading is that GLM's Vals result says something specific about the Flash variant in this harness, not something general about the GLM family.

Kimi K3.1 Pro, which we covered in our Kimi K3.1 Pro article, finished with 20 Lighthouse objectives, ahead of Grok 4.7 at 16 and Kimi K3 at 10. For wider context on how the newest frontier models compare on general benchmarks, see our GPT-6 Astra benchmark scores and our Cosmos 3 Nano vs Genie 4 comparison, and our model hub, Gemini Carbon vs Claude Opus 5.5 vs GPT-6 Astra.

Read this result carefully: four caveats

First, the networks were simulated. A cyber range is built to look like a real organisation, but it is not one, and nothing in this test shows what either model would do against a defended real target or says anything about real breach rates.

Second, Vals is the sole source for its own figures. The results were reported by WIRED and summarised by several outlets, but every number traces back to Vals' own leaderboard and methodology page. No independent lab has replicated the runs yet. The published figures also do not include run counts, score distributions or confidence intervals for every entry, so the eight-point gap should be read as Vals' reported result, not as a statistically precise measurement.

Third, the test rewards long, patient runs. DeepSeek made most of its progress after 100 million tokens, which means short evaluations understate it.

Fourth, the open model's advantage comes with no safety net. Vals notes that once a model's weights are public, no provider can monitor or stop how it is used.

Illustration: Close-up of hands typing on a laptop keyboard with a blurred glowing code screen

What this means for defenders

For security teams, the practical takeaway is not to fear the benchmark but to use it.

For the labs, the result weakens the case that restricted access keeps advanced cyber capability out of reach. It does not destroy that case: Daybreak still won, still worked far more efficiently, and the closed labs may hold stronger unreleased models.

Key Takeaways

GPT-6 Sol Daybreak led Vals CyberOffense at 80.8 percent of objectives, with open-weight DeepSeek V4.1 Flash 1017 close behind at 72.8 percent, a gap of about eight points.
DeepSeek matched Daybreak on the Nova network, averaging 16.3 of 21 objectives against 15.7, then trailed on the harder Lighthouse network, 21 of 31 against 27.
A DeepSeek run cost under 3 dollars against about 22 dollars for Daybreak, roughly a tenth of the cost, and WIRED put a full attack at 2 to 3 dollars versus 17 to 46 dollars.
DeepSeek scored nothing in its first 25 million tokens and made most progress after 100 million, so short compute-capped tests understate cheap open models.
No model completed either full attack chain, and Meta Wonder completed the most Lighthouse objectives, 28 of 31.
These were simulated networks, Vals is the sole source for its own figures, and the result says nothing about real breach rates.

Conclusion

DeepSeek V4.1 Flash did not beat GPT-6 Sol Daybreak, but it came within eight points on a test of autonomous network attacks, matched it on the first simulated network, and did it for under 3 dollars a run. That is the story: not a hacked corporation, but a closing gap between restricted and downloadable AI, measured on simulated ranges by a single benchmarking firm. Defenders should treat it as a cheap, practical invitation to test their own systems, and everyone else should treat it as a result to quote carefully, with its limits attached.

Search This Blog

INORYVA