Gemini 3.8 Flash: a million-token context window and a focus on agents

Gemini 3.8 Flash launched on September 2 as a fast model for programming, multi-step work, and autonomous agents. It accepts up to 1,048,576 input tokens and can return 65,536 tokens; text, images, audio, video, and PDFs are supported as input. Its documentation also plainly notes possible hallucinations and occasional slowdowns.

More than a fast chat model

The Flash label often suggests a tradeoff: speed in exchange for harder work. This release points elsewhere—long chains of actions. It supports function calling, file work, code execution, search, and preview computer use. That matters when an agent must read a specification, split a task into steps, edit a project, and then check the result.

Benchmark numbers need a little distance

The published evaluation mentions DeepSWE v1.1, Vals Finance Agent v2, Harvey’s Legal Agent Benchmark, and HLE-Verified. Those are useful signals, not a promise that an unattended model will handle your repository or contract correctly. The model card also says it may spend more tokens to improve an answer, and that knowledge freshness varies by domain.

A sensible first test

Try one small task on your own material: one document, one tool, and a crisp definition of done. Add actions only after that. A million-token context is handy, but a large attachment does not replace fact-checking or permission checks.

References: the model card and API documentation. For a working chat with public models, visit NeuralSpace Chat.