Vals benchmark puts Deepseek V4.1 Flash within 8 points of GPT-6 Sol on cyber offense
Vals AI's CyberOffense benchmark has narrowed the perceived divide between proprietary and open-weight models on simulated corporate network attacks. OpenAI's GPT-6 Sol Daybreak achieved 80.8 percent of attack objectives, while DeepSeek V4.1 Flash reached 72.8 percent within roughly eight points at about three dollars per evaluation run compared with roughly twenty-two dollars for the closed model, according to reporting that cites Vals. The results suggest capability parity is approaching faster than cost and access asymmetry alone would imply. Analysts quoted in the coverage warn that because open models cannot be tightly restricted, organizations may have less time to adapt defensive playbooks than policy debates sometimes assume. Readers should treat the scores as benchmark outcomes in simulated environments, not direct forecasts of real-world breach rates.
Vals benchmark puts Deepseek V4.1 Flash within 8 points of GPT-6 Sol on cyber offense
On Vals AI's CyberOffense benchmark, OpenAI's GPT-6 Sol Daybreak reached 80.8 percent of attack objectives in simulated corporate networks. The open-weight Deepseek V4.1 Flash hit 72.8 percent while costing about $3 per run versus roughly $22, The Decoder reports citing Vals.
Key takeaway
Open-weight DeepSeek V4.1 Flash now lands within eight points of GPT-6 Sol on Vals CyberOffense while costing roughly one-seventh per run, eroding the idea that only closed models pose elite offensive automation risk.
What happened
The Decoder reports, citing Vals AI, that on the CyberOffense benchmark OpenAI's GPT-6 Sol Daybreak reached 80.8 percent of attack objectives in simulated corporate networks, leading the cited comparison.
The same coverage attributes 72.8 percent to the open-weight DeepSeek V4.1 Flash, with per-run cost cited at about three dollars versus roughly twenty-two dollars for GPT-6 Sol Daybreak, and quotes analysts saying defenders face a shrinking window because open models can no longer be restricted.
Evidence
GPT-6 Sol Daybreak scored 80.8 percent of attack objectives on Vals AI's CyberOffense benchmark in simulated corporate networks.
The Decoder · attributed
On Vals AI's CyberOffense benchmark, OpenAI's GPT-6 Sol Daybreak reached 80.8 percent of attack objectives in simulated corporate networks.
DeepSeek V4.1 Flash scored 72.8 percent as an open-weight model on the same benchmark.
The Decoder · attributed
The open-weight Deepseek V4.1 Flash hit 72.8 percent while costing about $3 per run versus roughly $22, The Decoder reports citing Vals.
Analysts cited in the coverage say unrestricted open models shrink the time available for defenders.
The Decoder · attributed
Because open models can no longer be restricted, analysts say the window for defenders is shrinking fast.
Why it matters
Security teams must plan as if high-tier attack automation will diffuse through cheap, downloadable weights, not only through gated APIs whose use can be monitored or throttled.
Limits and uncertainties
Scores reflect Vals CyberOffense simulations in corporate network settings, not verified rates of successful intrusions against live enterprises.
The packet provides headline percentages and run costs via The Decoder citing Vals, without full benchmark methodology or task definitions.
Practical implications
Red-team and detection engineering should evaluate defenses against capable open-weight agent stacks, not assumptions that only closed frontier APIs reach offensive parity.
Risk and governance workflows should decouple model access controls from capability estimates when prioritizing incident readiness and vendor monitoring.
What to watch
Whether the next Vals CyberOffense leaderboard updates show additional open models clustering near or above the low seventies on objective completion.
If major vendors or CERT bodies publish defensive guidance or countermeasures explicitly tied to the same CyberOffense task suite.
