Google DeepMind launched Gemini 3.8 Flash on September 2, 2026, as an evolution of Gemini 3.7 Flash. The model is generally available and targets coding, agents, and complex knowledge workflows, with multimodal input and text output.
An evolution of the Gemini 3 family
The new version retains the technical foundation of Gemini 3.7 Flash. Google refers readers to the earlier model card for architecture, training, hardware, and software details, so it should not be described as a complete rebuild.
Google presents the model for users, developers, and organizations that need to carry out multi-step tasks. Its intended uses include software engineering, agent workflows, and the processing of extensive information.
Text, images, audio, and video as input
Gemini 3.8 Flash accepts text, images, audio, video, and PDF documents. Its documented output is text, which means multimodal understanding should not be mistaken for native image, audio, or video generation.
The context window reaches up to one million tokens, while maximum output is documented at 64,000 tokens. These capacities support lengthy inputs, but they do not guarantee that every detail will be interpreted or preserved correctly.
Tools for agents and coding
The documentation confirms function calling, search as a tool, and computer use. These features can support workflows in which the model retrieves information, uses external services, or performs actions under rules established by an application.
Configurable effort levels provide another way to balance quality, cost, and latency. Google warns that more demanding settings may consume additional tokens without ensuring that every increase will produce a correct response.
Availability across multiple products
Gemini 3.8 Flash is distributed through the Gemini app, Google AI Studio, Gemini API, Gemini Enterprise Agent Platform, AI Mode, and Google Antigravity. Google lists the model itself as generally available.
Actual access may vary by product, plan, and region. The available documentation does not establish that every user receives identical tools, quotas, or commercial terms, or that the service is free or unlimited.
Results published by Google
Google reports a 54.9% result on HLE-Verified and improvements over Gemini 3.7 Flash in selected coding, finance, and legal evaluations. These figures come from tests published by the company and do not demonstrate universal superiority.
Official demonstrations include interactive applications and long-running development tasks. They are selected examples of possible capabilities, not a promise that every project will achieve the same quality, speed, or stability.
Limits that require supervision
The model card acknowledges possible hallucinations, latency, and timeouts. It also gives a general knowledge cutoff of March 2026, while warning that knowledge in some areas may remain limited to January 2025.
Google also reported a slight regression in multilingual safety compared with Gemini 3.7 Flash in automated evaluations. For these reasons, responses, code, and agent actions should be reviewed before they are used in important processes.
