このモデルの強みは何ですか？

3.8 Flashの1Mコンテキスト・マルチモーダル基盤をセキュリティに特化自律脆弱性発見・自動パッチ適用にチューニング Fairwindで政府・重要インフラ防御者のみに開放、責任ある配布 3.8 Flashと同じ$0.75/$3.75で導入

このモデルの弱みは何ですか？

一般提供せず、Fairwindの審査通過が前提で実戦投入までハードルベースは3.8 Flashのため最高難度の汎用エージェント性能は上位モデルに劣る防御用途に特化のため汎用タスクへの流用は期待できないアクセス制限ゆえに独立ベンチマークがほぼ存在しない

どんな用途に最適ですか？

政府・重要インフラ向けの自律脆弱性スキャンと優先順位付けパッチ提案・自動修正のドラフト生成既存セキュリティパイプラインへのエージェント組み込み 3.8 Flash APIと同じインフラでの防御運用

Google DeepMindプロプライエタリ

Gemini 3.8 Flash Cyber

Name: Gemini 3.8 Flash Cyber
Author: Google DeepMind

このモデルを比較

Gemini 3.8 Flash Cyberは、Google DeepMindが2026年9月2日にベースのFlashモデルと同時に発表したセキュリティ特化型の派生モデル。同じ105万トークン・コンテキストとマルチモーダル基盤を持ち、自律的な脆弱性発見と自動修正（パッチ適用）にチューニングされている。アクセスはFairwindプログラム経由にゲートされ、政府および重要インフラの防御者に限定され一般提供はされていない。価格は3.8 Flashと同じ100万トークンあたり$0.75/$3.75。

シェア:X はてブ LINE

パラメータ

非公開

コンテキスト長

ライセンス

プロプライエタリ

リリース日

2026-09-02

API料金

入力料金（1Mトークンあたり）

$0.75

出力料金（1Mトークンあたり）

$3.75

課金モード: standard

強み

•3.8 Flashの1Mコンテキスト・マルチモーダル基盤をセキュリティに特化
•自律脆弱性発見・自動パッチ適用にチューニング
•Fairwindで政府・重要インフラ防御者のみに開放、責任ある配布
•3.8 Flashと同じ$0.75/$3.75で導入

弱み

•一般提供せず、Fairwindの審査通過が前提で実戦投入までハードル
•ベースは3.8 Flashのため最高難度の汎用エージェント性能は上位モデルに劣る
•防御用途に特化のため汎用タスクへの流用は期待できない
•アクセス制限ゆえに独立ベンチマークがほぼ存在しない

活用例

•政府・重要インフラ向けの自律脆弱性スキャンと優先順位付け
•パッチ提案・自動修正のドラフト生成
•既存セキュリティパイプラインへのエージェント組み込み
•3.8 Flash APIと同じインフラでの防御運用

深度分析

Base model

Gemini 3.8 Flash

GA September 2, 2026; 1,048,576-token context, 65,536 output

Focus

Autonomous vulnerability discovery & automated patching

Defensive cybersecurity, not general coding

Access

Fairwind Program

Governments and critical-infrastructure operators only

Pricing

$0.75 / $3.75 per 1M

Shared with 3.8 Flash; doubles January 1, 2027

CyberGym

Frontier-level (vendor-claimed)

Vulnerability-discovery benchmark; independent replication pending

強み

・Google's first model purpose-built for defensive cyber work, not just general-purpose coding
・Gated behind the Fairwind Program, keeping the most dangerous capability out of the open API
・Inherits 3.8 Flash's cost-efficient agentic coding (Terminal-Bench 2.1 90.8%) and 1M-token context

弱み

・Reachable only through a vetting program — most teams cannot use it yet
・Cyber-capability benchmarks remain vendor-reported; independent replication is outstanding
・Shares 3.8 Flash's January 1, 2027 price doubling, which can raise real cost ~40%

競合比較

Model	Arena	Price
GPT-6 Astra	Critical cyber tier	$10/$50
Claude Mythos 5.1	Cyber via Claude Security (invitation)	$10/$50
Gemini 3.8 Flash Cyber	Fairwind Program	$0.75/$3.75

概要

Gemini 3.8 Flash Cyber, released September 2, 2026 alongside Gemini 3.8 Flash, is Google DeepMind's first model built specifically for defensive cybersecurity — autonomous vulnerability discovery and automated patching — rather than general-purpose coding. Google is making it available only through the new Fairwind Program to governments and critical-infrastructure operators, a gating pattern that mirrors the trusted-access tiers OpenAI and Anthropic used for their own high-capability cyber models the same week. It inherits 3.8 Flash's economical agentic coding (Terminal-Bench 2.1 90.8%) and 1,048,576-token context, priced at the same $0.75/$3.75 per million tokens (doubling January 1, 2027). Google reports frontier-level results on the CyberGym vulnerability-discovery benchmark; independent verification is outstanding. The move signals that the three leading US labs now treat offensive-grade cyber skill as too sensitive for open release.

ベンチマーク＆性能

As a derivative of Gemini 3.8 Flash, 3.8 Flash Cyber carries the base model's published coding numbers — 73.7% on DeepSWE v1.1, 90.8% on Terminal-Bench 2.1 — and layers domain tuning for security workflows. Google states frontier-level performance on CyberGym for vulnerability discovery; exact figures and methodology have not been independently replicated. No public HLE, GPQA, or general-reasoning scores are separated out for the Cyber variant.

詳細比較

Against GPT-6 Astra (OpenAI's 'Critical' cyber tier, $10/$50, trusted access + Daybreak) and Claude Mythos 5.1 (Anthropic's permissive-safeguard variant behind a trusted-access program, $10/$50, powering Claude Security), Gemini 3.8 Flash Cyber is the cheapest to run at $0.75/$3.75 but the most narrowly scoped — defensive cyber only, not a general agent. All three share the same week's theme: the strongest cyber capability ships gated, not open. Astra and Mythos reach broader enterprise via paid tiers; 3.8 Flash Cyber is restricted to vetted government and infrastructure partners.

コミュニティ評価

Reaction among security practitioners has been cautious-positive: the gating is read as responsible, but the lack of open access means independent red-teaming will be slow. Some researchers noted the irony that three frontier labs all chose to withhold offensive cyber skill within days of each other, suggesting regulatory or internal-risk pressure rather than purely technical limits. The Fairwind Program's eligibility criteria were not fully detailed at launch.

ユースケース

1. Defensive vulnerability discovery — continuously scan first-party codebases and infrastructure for unknown flaws, prioritizing by exploitability. 2. Automated patching — generate and test fixes for discovered vulnerabilities inside isolated pipelines. 3. Critical-infrastructure protection — government and operator teams use it under the Fairwind Program for threat modeling and remediation at scale. Teams outside the program should treat it as unavailable and plan around 3.8 Flash for general coding.