Six weeks ago Google refreshed its Flash line, and the company is already back with another. Gemini 3.8 Flash landed September 2 as the third Flash generation in that span, carrying what DeepMind calls its best reasoning and coding model yet at roughly the speed and price of its predecessor.
The launch is really two products sharing one brain. Most developers get Gemini 3.8 Flash with the standard safety stack. Security teams get Gemini 3.8 Flash Cyber, a sibling with looser cyber safeguards that DeepMind only hands to vetted defenders inside its Fairwind program. Training time spent in demanding cybersecurity work sharpened both.
On long-horizon coding, DeepMind says 3.8 Flash beats most larger frontier models on DeepSWE v1.1 at a fraction of the cost, and it posts a 54.9 percent score on HLE-Verified. The Cyber variant reports frontier-level results on the CyberGym vulnerability benchmark, clears a 70 percent success bar on an internal 20-language test, and reaches 47.2 percent pass@1 on Collinear’s CWE-Bench patching challenge, within a point of the 47.8 percent leader. DeepMind prioritized fixing flaws over weaponizing them, and Google already runs the model over its own codebases.
Safeguard policy separates the pair. The open model keeps CBRN and cyber-offense mitigations in place; the Cyber edition ships with more permissive rules and exists solely to arm defenders. DeepMind also reports markedly better resistance to prompt injection attacks in third-party Gray Swan testing.
Discounted launch pricing ends December 31. From January 1, input tokens run $1.50 per million and output tokens $7.50 per million.
