Gemini 3.5 Flash

The First Gemini 3.5 Model

Google introduced Gemini 3.5 Flash on May 19, 2026 as the first model in the Gemini 3.5 family.

The launch theme was frontier intelligence with action.

That positioning matters because Gemini 3.5 Flash was not introduced as a lightweight fallback for simple requests. Google explicitly targeted difficult coding, autonomous agents, multi-step workflows, sub-agent deployment, and long-horizon tasks.

Google also reported that the model outperformed Gemini 3.1 Pro on several coding and agentic benchmarks while running substantially faster than other frontier models.

The stable Gemini API model ID is gemini-3.5-flash.

It accepts text, images, video, audio, and PDFs and generates text. The model supports up to 1,048,576 input tokens and up to 65,536 output tokens.

Gemini 3.5 Flash also supports dynamic thinking, automatic thought preservation, Google’s built-in tool ecosystem, context caching, and Computer Use in preview.

• Stable model ID: gemini-3.5-flash.
• Released May 19, 2026.
• Stable and GA for scaled production use.
• 1,048,576 input tokens.
• 65,536 maximum output tokens.
• Text, image, video, audio, and PDF input.
• Text output.
• Default thinking level: medium.

Coding Quality Is a Trajectory Property

Repository-level engineering rarely succeeds in one generation.

A coding agent may need to:

1. inspect source files;
2. understand dependencies;
3. form a plan;
4. modify several files;
5. execute tests;
6. diagnose failures;
7. search for more context;
8. repair the implementation;
9. verify the final state.

Gemini 3.5 Flash was explicitly built for this loop.

Google’s launch materials emphasized coding, sub-agent deployment, long-horizon task execution, and rapid iteration over alternate approaches.

For production evaluation, measure more than code syntax.

Useful metrics include:

• accepted implementation rate;
• first-pass test success;
• number of patch iterations;
• repeated file reads;
• tool calls;
• failed commands;
• unnecessary edits;
• human intervention;
• total tokens to acceptance.

This makes 3.5 Flash particularly valuable as a historical baseline for later Gemini Flash generations, which progressively focused on reducing reasoning steps, output tokens, and tool-call overhead.

From Tool Calls to Interface Actions

Before this integration, Google exposed Computer Use through a specialized model.

With Gemini 3.5 Flash, Computer Use became a built-in tool in the primary Flash model.

This is important for agents that need to act in systems without a clean API.

A computer-use agent can:

• inspect the current screen;
• reason about UI state;
• click controls;
• type into fields;
• navigate between pages;
• observe the updated interface;
• continue until the task is complete.

Google positioned this capability for browser automation, enterprise applications, continuous software testing, and knowledge-work workflows.

Computer Use remains a preview capability and requires stronger production safeguards than ordinary text generation.

Evaluation should include wrong clicks, destructive actions, navigation loops, state recovery, confirmation gates, and successful task termination.

Long Context Can Hold the Whole Working Set

For an agent, context is not just the user’s prompt.

It can include:

• repository files;
• screenshots;
• long PDFs;
• video;
• audio;
• tool outputs;
• retrieved URLs;
• system instructions;
• prior conversation turns;
• preserved reasoning state.

Gemini 3.5 Flash can reason across these modalities in one model.

That is useful for long-horizon coding, multimodal research, application automation, document workflows, and agents that need substantial state.

But capacity is not the same as relevance.

Sending every available file, turn, and tool result can increase cost and make important instructions harder to identify.

Use retrieval, caching, selective history, media-resolution controls, and context pruning to keep the working set useful.

Combined Tool Use Matters for Agents

Google lists support for:

• Google Search grounding;
• Google Maps grounding;
• File Search;
• code execution;
• URL context;
• function calling;
• structured outputs;
• context caching;
• Computer Use in preview;
• Batch API;
• Flex inference;
• Priority inference.

Gemini 3 also supports combined tool use.

An agent can, for example, use Search for current evidence, inspect a URL, execute code to calculate a result, call an application-defined function, and return a structured response.

This makes tool orchestration itself an evaluation target.

Measure wrong-tool calls, unnecessary calls, invalid arguments, tool-result interpretation, recovery from failures, and whether the model knows when to stop.