Skip to content
NewsModel release

Google Releases Gemini 3.8 Flash and a Restricted Cyber Variant for Vulnerability Detection

· by Pondero Newsdesk

The short version

Google's fourth Flash-tier model in four months scores 59 on the Artificial Analysis Intelligence Index and ships with a purpose-built cybersecurity sibling locked to vetted Fairwind Program partners.

Google Releases Gemini 3.8 Flash and a Restricted Cyber Variant for Vulnerability Detection

Google released two models on September 2, 2026: Gemini 3.8 Flash, its fourth Flash-tier model in four months, and Gemini 3.8 Flash Cyber, a restricted variant purpose-built for software vulnerability detection that requires separate approval through the Fairwind Program to access.

What

Gemini 3.8 Flash is available through the generateContent API, Google AI Studio, Android Studio, Stitch, and the Gemini Enterprise Agent Platform. It scores 59 on the Artificial Analysis Intelligence Index, a 3-point gain over Gemini 3.7 Flash, tying it with GPT-5.6 Sol and Grok 4.6 at high-reasoning settings. Google highlights agentic and multi-step reasoning tasks as the model's primary improvement area, with a 12-point gain on banking tasks in the Vals Finance Agent V2 benchmark and a 54.9% score on Humanity's Last Exam. Pricing during the introductory period is $0.75 per million input tokens and $3.75 per million output tokens, per VentureBeat; the standard post-promotional rate is $1.50 per million input and $7.50 per million output, per Artificial Analysis. The 1-million-token context window carries forward from prior Flash releases.

Gemini 3.8 Flash Cyber ships exclusively through the Fairwind Program, restricted to governments, critical infrastructure operators, and vetted software maintainers. Per VentureBeat, it scores 86.2% on the CyberGym benchmark and 47.2% on CWE-Bench. In internal Google testing it produced 2.6 times more correct Chrome vulnerability patches than larger commercial models. Third-party testing by Wiz showed 7.5% to 9.7% higher recall at 2.3 to 5.2 times lower cost per detection. Standard API pricing does not apply; access requires application through the Fairwind Program.

Why it matters

The cost gap is concrete. Artificial Analysis puts Gemini 3.8 Flash at $0.58 per Intelligence Index task, making it the cheapest model at its intelligence tier. The Register reports Claude Fable 5.1 costs roughly $3.76 per equivalent task. Teams routing complex agentic workloads through higher-priced models should run a direct cost-per-task comparison against Gemini 3.8 Flash. Note that the model uses higher token counts per output than its predecessor, which partially offsets the lower per-token rate; the $0.58 figure already accounts for that.

For security teams: Flash Cyber's Fairwind Program approval model mirrors the gated-access approach OpenAI uses for its Astra Daybreak Blue program. Restricted-access dual-use security AI appears to be the industry's emerging standard. Organizations that want early access to Flash Cyber should apply through the Fairwind Program now rather than assume broader availability follows.

Context

The four-Flash-in-four-months cadence responds to sustained competitive pressure. The Register notes that Google faced questions about its frontier model commitment after leadership changes and delayed Gemini 3.5 Pro timelines. Gemini 3.8 Flash's 59 Intelligence Index score trails Claude Opus 5 (63) and Claude Fable 5.1 (66), so the standard model does not yet lead the frontier at the highest difficulty tasks. Google ranks the model No. 14 in Arena.ai's Agent Arena and No. 7 in Text Arena.

What to watch next

Google has not announced a Gemini 3.8 Flash Pro follow-on, but the four-month cadence points toward a potential October release. A second question: whether Google publishes GPAI-compliant capability evaluations for Flash Cyber to match the methodology used in OpenAI's Preparedness Framework disclosures.

Sources