Vals AI: OpenAI's GPT-6 Sol Daybreak hits 80.8% on CyberOffense, open-weight Deepseek V4.1 Flash reaches 72.8%

OpenAI's GPT-6 Sol, part of its restricted Daybreak cyber program, posted the best result on the Vals AI CyberOffense benchmark at 80.8 percent of objectives. Deepseek V4.1 Flash 1017, a freely available open-weight model, reached 72.8 percent and costs about $3 per run instead of roughly $22, according to Vals AI, as reported by The Decoder.

How the benchmark works

Vals CyberOffense simulates realistic corporate networks. Each agent works alone on an isolated Kali Linux attack machine, starting with only a public web address and no human help. The agent sends commands to the simulated network on its own and works its way from the web application through Windows servers to admin privileges. Every action is logged externally and scored automatically.

Results

The benchmark includes two scenarios: Nova (21 objectives) and Lighthouse (31 objectives).

Model	Nova (of 21)	Lighthouse (of 31)
GPT-6 Sol Daybreak	15.7 (average of 3, range 12-19)	27
Deepseek V4.1 Flash 1017	16.3 (average of 3, range 15-18)	21
Kimi K3.1 Pro	13	20
Meta Wonder	7	28
Grok 4.7	12	16
Kimi K3	9	10
GLM 5.3 Flash	5	5

In Lighthouse, Daybreak took over admin control of the first company, stole passwords, and used them to break into the second. It completed 27 of 31 objectives in under three hours. Deepseek completed 21 objectives at a tenth of the cost.

Token budget matters

The more tokens an agent uses, the more objectives it reaches. Most models do not make meaningful progress until they pass 100 million tokens. Deepseek scored zero results in its first 25 million tokens and made most of its progress after 100 million. Vals concludes that short tests underestimate the real potential of open models, and that this applies to most open Chinese models.

Vals researchers also say that once a model's weights are public, no provider can control or restrict how it is used.

Earlier findings on the open-closed gap

The Decoder links the results to a July analysis by the UK's AI Security Institute (AISI). It found the gap between open and closed frontier models in cyber capabilities has shrunk to four to seven months. For most of 2025, the gap was six to ten months.

AISI also reported a large cost difference. A cyber-range run with Deepseek V4-Pro costs $1.19, compared with $85 for Anthropic's Opus 4.5. The safety measures on the open models proved largely ineffective, according to the report.

Separately, a recent Anthropic study showed that the freely available GLM-5-3 can build exploits at nearly the same level as Anthropic's internal top model, Mythos.

What this means

The Vals data puts an open-weight model within reach of the top restricted model on Nova, where the Deepseek average (16.3) is slightly above the Daybreak average (15.7), while the per-run cost is about $3 versus roughly $22. Lighthouse shows Daybreak ahead of Deepseek, 27 objectives to 21, and Meta Wonder's 28 is the highest count in the table.

The token-budget finding affects how results should be read: Deepseek produced nothing in its first 25 million tokens, so evaluations with short runtimes may understate what open models can do. Because weights are public, restrictions by the provider cannot be enforced after release, as Vals researchers note.

Vals also sees an upside for defenders: the same models that exploit vulnerabilities can help companies find and fix them first. Vals expects attacks on real networks to increase in the coming months.