# Gemini 3.8 Flash Brings Cheap Reasoning to Cyber
_Google debuts Gemini 3.8 Flash and a Cyber variant for defenders, pairing frontier patching scores with Flash pricing via the Fairwind Program._
**Published:** 2026-09-02
**Source:** https://www.startuphub.ai/ai-news/ai-research/2026/gemini-3-8-flash-brings-cheap-reasoning-to-cyber
---
[Gemini 3.8 Flash](https://deepmind.google/blog/introducing-gemini-3-8-flash-and-38-flash-cyber/) arrived Sep 2 as Google DeepMind's third Flash in six weeks, holding 3.7 Flash pricing at $0.75 input and $3.75 output per million tokens until Dec 31.

Tulsee Doshi, senior director of product management, and Raluca Ada Popa, Gemini Security Lead at [DeepMind](https://www.startuphub.ai/ai-news/ai-research/2026/deepmind-s-game-ai-evolves), frame both variants as one core intelligence split between a general workhorse and a gated Cyber model for trusted defenders via the Fairwind Program, according to [DeepMind](https://deepmind.google/blog/introducing-gemini-3-8-flash-and-38-flash-cyber/).

The demo targets code security across C and C++ CyberGym tasks and an internal 20-language benchmark, plus Chrome and Google Cloud codebases, with no attacker access required but defender vetting required for Cyber.

## How Gemini 3.8 Flash actually works

Google says the gains came from long-running agentic loops that recursively evaluate and refine the model, paired with rigorous training in cybersecurity to harden reasoning.

At inference the model trades efficiency for diligence. It takes extra reasoning steps and iterative tool calls, burning more tokens at higher effort levels, while lower effort or staying on 3.7 Flash preserves cost.

Think of it like a tireless junior engineer who reruns tests and re-reads docs until the patch holds, not a faster autocomplete that stops at the first plausible answer.

## Why this matters, and what it does not fix

At Flash cost it nears frontier results, with 54.9 percent on HLE-Verified, wins on DeepSWE v1.1 long-horizon engineering, plus Vals Finance Agent V2 and the Harvey Legal Agent Benchmark. For security, it posts 47.2 percent pass at 1 on CWE-Bench versus 47.8 percent for a larger frontier model, over 70 percent success across 20 languages, 2.6 times more correct Chrome patches than larger commercial models, and 7.5 to 9.7 percent higher recall for [Wiz](/startups/wiz) at 2.3 to 5.2 times lower cost.

That cost curve matters for defenders who need speed, and Google reports a real Cloud Vulnerability Research find in under two hours, plus prompt injection robustness gains measured by Gray Swan. Yet Cyber remains gated, and the base model keeps CBRN and cyber offense safeguards per the Frontier Safety Framework.

What isn't fixed is access and overhead. Introductory pricing doubles on Jan 1, 2027 to $1.50 and $7.50, high effort inflates tokens, and Google hasn't disclosed false positive rates or independent reproductions beyond its chosen benchmarks, so builders should budget for verification and keep 3.7 Flash in reserve for efficiency-first paths.
---
Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.