DeepSeek V4.1 Flash Tops the CyberGym Security Benchmark

DeepSeek's newest lightweight model just took the top spot on a security benchmark that is unusually hard to game. The scores come from official data DeepSeek shared, and they put a Flash-tier model ahead of the flagships.

DeepSeek-V4.1-Flash scored 88.1% on CyberGym, with Opus 5 and GLM-5.3 tied behind it at 84.5% each and Kimi-K3 at 80.0%. It did that at Flash pricing, where a run costs a fraction of a cent.

What Happened

CyberGym is built by UC Berkeley, and it works nothing like a synthetic coding test. Agents are dropped into real codebases carrying more than 1,500 historical vulnerabilities, drawn from projects like OpenSSL and FFmpeg.

To score, an agent has to find the flaw, write a working proof-of-concept exploit, and ship a patch. There is no partial credit, and a hallucinated pass does not count.

On CyberGym the standings ran:

• DeepSeek-V4.1-Flash — 88.1%
• Opus 5 — 84.5%
• GLM-5.3 — 84.5%
• Kimi-K3 — 80.0%

The same model showed up strong on other agentic evals. On DeepSWE v1.1 it hit 74.2%, just ahead of Opus 5 at 74.0% and GPT-5.6-Sol at 73.0%. On Automation-Bench it reached 54.8%, beating Opus 5 at 50.3% and GLM-5.3 at 48.8%.

One result went the other way. On Terminal Bench 3.0 it landed at 30.0%, second to Opus 5 at 43.3%.

Why It Matters

An 88.1% success rate on hardened security pipelines at a sub-cent price changes what is affordable. Automated red-teaming and repo auditing stop being a budget line only large teams can carry, and start being something you can run at scale.

The caveat is Terminal Bench 3.0, where the Flash model trails Opus 5 by more than 13 points. Strong on vulnerability work does not mean strong everywhere.