4 September 2026

Google releases Gemini 3.8 Flash, a cybersecurity variant

First reported

Ars Technica ran this on , 2 days before the other 3 sources picked it up.

  • Google released Gemini 3.8 Flash, a general-purpose model priced at introductory rates through year-end: $0.75 per million input tokens and $3.75 per million output tokens.
  • Gemini 3.8 Flash Cyber, a specialized variant tuned for finding and fixing security vulnerabilities, currently available only to governments and trusted testers, showed 2.6x better patch accuracy than prior versions.
  • On DeepSWE, a benchmark measuring complex software engineering problem-solving, Gemini 3.8 Flash ranked highest while costing less than competitors, but trails Claude Opus significantly on computer-use tasks.
  • Gemini 3.8 Flash can organize and edit Google Photos based on text prompts, preserving originals and requesting user confirmation before sharing.

Where they differ

  • The Neuron

    focused exclusively on a new Photos feature the other newsletters did not mention; all three agreed on pricing and general capability claims.

  • Deep Learning Weekly

    both emphasized the cybersecurity variant's technical improvements.

What each one reported

AI BreakfastIndependent editors

Google released Gemini 3.8 Flash at competitive pricing with strong DeepSWE performance, and a gated 3.8 Flash Cyber model accessible only through the Fairwind Program for national cyber authorities and critical infrastructure. The newsletter notes frontier benchmark scores do not match frontier real-world findings, citing startup AISLE finding six curl CVEs that leading models missed.

The NeuronPete Huang & Grant Harvey

Google's Gemini Spark gained capabilities to find, enhance, organize, and prepare photos from a single prompt while keeping originals safe and requiring user confirmation before sharing.

Deep Learning WeeklyEditorial team

Google released Gemini 3.8 Flash and a cyber-specialized variant that produces 2.6x more correct vulnerability patches than competing commercial models.

Reported by Ars Technica