Google Launches Gemini 3.8 Flash and Access-Gated Cyber Variant

Google DeepMind announced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026, arriving three weeks after the release of Gemini 3.7 Flash. [1][5] This launch marks the company's third Flash-tier model release in a six-week period. [1][4] Google made Gemini 3.8 Flash generally available immediately through the Gemini API, Google AI Studio, Android Studio, Antigravity, and Gemini Enterprise. [1][5]

Developers can route production traffic to the general model using the `gemini-3.8-flash` API identifier. [5] The general model is available for an introductory price of $0.75 per million input tokens and $3.75 per million output tokens until December 31. [3] Both variants run on the same foundational intelligence and remain API-only, meaning there is no self-hosted or on-premises deployment path because Google has not published the model weights. [1][5]

The Gemini 3.8 Flash architecture maintains the same specifications as 3.7 Flash, featuring a 1,048,576-token context window and a 65,536-token maximum output. [1] The model accepts text, image, audio, and video input while producing text output. [1] The research team notes that 3.8 Flash works harder than previous versions by executing extra reasoning steps on complex tasks. [1]

The releases are powered by shared foundational intelligence and further accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models. [4] Google maintained the Gemini API's thinking levels at low, medium, and high parameters, with medium remaining the default. [1] The minimal thinking level is no longer supported on 3.8 Flash, and setting it will return an API validation error. [1]

While the general Flash model targets software engineering and autonomous agents, Gemini 3.8 Flash Cyber is built specifically to find software vulnerabilities and write patches for them. [4][5] Google claims the Cyber variant achieves frontier-level performance in vulnerability detection and automated patching. [4] Google states the shared model core gained significant coding and reasoning improvements through rigorous training in the cybersecurity domain. [4]

Unlike the general model, Gemini 3.8 Flash Cyber is not openly deployable. [1][5] Google gates access to this model on a case-by-case basis through a new initiative called the Fairwind Program, which is aimed at trusted defenders. [4][5] Maintaining an API-only approach keeps the model's vulnerability-hunting logic contained within Google's infrastructure. [5]

Google's bifurcated release strategy highlights a shift in how frontier AI developers handle specialized, high-risk capabilities. Instead of releasing a single model for the public to adapt, Google is reserving its vulnerability-hunting logic for vetted defenders while offering the broader reasoning benefits to general developers. By gating 3.8 Flash Cyber through the Fairwind Program and keeping all model weights closed, Google ensures its automated patching systems remain securely on its own infrastructure. Google claims the standard 3.8 Flash model beats rival systems like Claude Opus 5 and GPT-5.6 Sol on some benchmarks, positioning its intermediate tier against competitors' flagship models. What the sources don't address: What specific criteria Google will use to vet applicants for the Fairwind Program.