For Developers/Models/Gemini 3.8 Flash Cyber
Closed SourcePreviewGoogleReleased 2026-09

Gemini 3.8 Flash Cyber

Google's restricted Flash model for finding and patching software vulnerabilities

Context window

Not announced

Input / 1M tokens

Not announced

Output / 1M tokens

Not announced

Provider

Google

Google has not published pricing or a context window, and the model is not on the Gemini API pricing page. Its launch post says the model keeps Flash speed and cost, and reports a CWE-Bench score close to a leading frontier model at significantly lower cost. · Data verified 2026-11-11

Gemini 3.8 Flash Cyber (September 2, 2026) is a cybersecurity version of Gemini 3.8 Flash, built on the same base model and tuned for vulnerability discovery and automated patching. It ships with more permissive cybersecurity mitigations than the general model, so Google limits it to trusted defenders: government authorities, critical infrastructure operators and software maintainers who apply through its Fairwind Program, where it can run with Google DeepMind's CodeMender agent harness. Google says it beats Gemini 3.5 Flash Cyber and larger frontier models on CyberGym, and that Chrome's security team got 2.6 times more correct vulnerability patches from it than from the best commercial models. It replaces Gemini 3.5 Flash Cyber.

How to access it

Restricted. Trusted government authorities, critical infrastructure operators and software maintainers apply for access through Google's Fairwind Program (deepmind.google/fairwind-program). Google has not published an API model ID or opened self-serve access.

Strengths

  • ✓86.2% on CyberGym, ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%) in Google's comparison
  • ✓Built on Gemini 3.8 Flash, so it keeps Flash-class speed and cost
  • ✓Works with the CodeMender harness for finding, triaging and patching vulnerabilities
  • ✓Chrome security team reports 2.6 times more correct patches than the best commercial models (Google-reported)

Best for developers who...

Vulnerability discovery across large codebasesAutomated patch generationSecurity teams in government, critical infrastructure and open-source maintenance

When to choose it (and when not to)

Reach for Gemini 3.8 Flash Cyber when...

  • →Your organization qualifies for the Fairwind Program and needs high-volume vulnerability scanning and patching
  • →You want a faster, cheaper cyber model than frontier-size options for continuous scanning

Look elsewhere if...

  • ✕You need public, self-serve API access
  • ✕You need Google's strongest cyber results: on Google's Fairwind page Gemini 4 Argon scores higher (85.8% vs 71.0% on real-world vulnerability discovery)
  • ✕General coding or chat work (use Gemini 3.8 Flash)

Benchmarks

BenchmarkScoreNotes
CyberGym86.2%Pass@1, from the chart in Google's launch post as reported by SiliconANGLE; Claude Mythos 5 83.8%, GPT-5.6 Sol 83.6%
CWE-Bench, Collinear run47.2%Pass@1 run by Collinear and reported by Google, vs 47.8% for an unnamed leading frontier model
Real-world vulnerability discovery (Google internal)71.0%Source code in 20 languages, from Google's Fairwind Program page; Gemini 4 Argon 85.8%
Wiz penetration testing benchmark58.2%From Google's Fairwind Program page; Gemini 4 Argon 70.9%

Source: Google launch post: Gemini 3.8 Flash and 3.8 Flash Cyber

Compare Gemini 3.8 Flash Cyber

Compare Gemini 3.8 Flash Cyber with any other model

Build a comparison →
All model comparisons →

Learn the concepts