Vals AI: DeepSeek V4.1 Flash edged GPT-6 Sol in one preliminary simulated network attack test

The evaluation company Vals AI says DeepSeek V4.1 Flash 1017 “beats GPT 6 Sol on our multi-step offensive cybersecurity task”. The result comes from a benchmark Vals is still building, set in simulated corporate networks, and Vals calls it a “first immediate finding” ahead of publishing the full evaluation “in the coming weeks”. GPT-6 Sol was tested through OpenAI’s Daybreak trusted access program.

What the chart shows

Vals’ chart counts objectives completed out of 21 against tokens spent:

• DeepSeek V4.1 Flash 1017: 16.3 on average over three runs.
• GPT-6 Sol (Daybreak): 15.7 on average over three runs.
• Single runs for the rest: Grok 4.7 12, Kimi K3 9, GLM 5.3 Flash 5, and Claude Opus 5.5 1, marked as having refused partway through the run.

So the lead is 0.6 objectives on average, in one scenario. The same chart shows DeepSeek spending several times as many tokens as Sol to get there. Vals has not published the scenario, the scoring rules or the run logs, so nobody can check the result yet.

The public number points the other way

This is a different test from Vals’ public CyberBench v1.1, which asks agents to reproduce and patch known open-source crash bugs. There, GPT-6 Sol scores 77.98% and DeepSeek V4.1 Flash 73.69%, at $1.34 against $0.07 per test. Both results can be true: one measures finding and fixing a bug in a single project, the other moving through a network toward objectives.

Why it matters

The cheap model in this race is DeepSeek’s open-weight V4.1 Flash, which anyone can download, while Sol’s cyber configuration sits behind OpenAI’s vetted-access program. If the full benchmark holds the result up, that gap in who can use which model is the policy question, and it is why the details Vals has not yet published matter more than the headline.