Gemini 3 Flash Preview

The First Gemini 3 Flash

Google released gemini-3-flash-preview on December 17, 2025.

At launch, Google described the model as frontier intelligence built for speed.

The important product shift was that Flash was no longer positioned only as a cheaper model for straightforward requests. Gemini 3 Flash brought strong reasoning, multimodal understanding, coding, and agentic behavior into the faster tier.

Google reported that the model surpassed Gemini 2.5 Pro across many benchmarks while operating at lower cost and higher speed.

Today, Google’s pricing documentation describes Gemini 3 Flash Preview as a legacy Flash model providing baseline speed and intelligence.

The endpoint remains available in public preview and currently has no announced shutdown date, but Google’s deprecation table recommends gemini-3.6-flash as its replacement.

The model supports:

• 1,048,576 input tokens;
• 65,536 output tokens;
• text, image, audio, video, and document input;
• text output;
• dynamic thinking;
• structured outputs;
• context caching;
• search grounding;
• code execution;
• function calling;
• URL context;
• Computer Use in preview.

Coding Is More Than Generating a Function

Google brought Gemini 3 Flash into Gemini CLI immediately at launch.

The model was designed for workflows where the AI repeatedly:

• reads code;
• searches files;
• reasons about dependencies;
• edits implementations;
• runs commands;
• interprets failures;
• retries;
• validates the final result.

Google reported a 78% SWE-bench Verified score for Gemini 3 Flash in its December 2025 Gemini CLI announcement.

That benchmark number is useful context, but production evaluation should focus on complete engineering trajectories.

Measure:

• first-pass test success;
• number of patch iterations;
• repeated file reads;
• invalid tool calls;
• command failures;
• unwanted edits;
• human intervention;
• latency to accepted implementation.

This page therefore treats Gemini 3 Flash Preview as an agentic-coding baseline, not merely as a generic chat model.

Audio Input Has a Separate Rate

Google currently lists Standard pricing at:

• $0.50 per 1M text, image, or video input tokens;
• $1.00 per 1M audio input tokens;
• $3.00 per 1M output tokens, including thinking;
• $0.05 per 1M cached text, image, or video tokens;
• $0.10 per 1M cached audio tokens;
• $1.00 per 1M cached tokens per hour of storage.

Batch and Flex pricing reduce token rates to:

• $0.25 text/image/video input;
• $0.50 audio input;
• $1.50 output.

Priority pricing increases rates to:

• $0.90 text/image/video input;
• $1.80 audio input;
• $5.40 output.

The Gemini Developer API also provides a free tier for Gemini 3 Flash Preview standard token usage.

Thinking tokens are included in output billing, so high-default reasoning can make the effective cost larger than the visible final response suggests.

The Preview Was Tuned Toward Reasoning Depth

Google’s Gemini 3 developer guide documents:

• minimal;
• low;
• medium;
• high.

high is the default and uses dynamic thinking.

That means an application that does not explicitly configure reasoning may spend more time and output tokens thinking than a later Flash integration that defaults to medium.

Use:

• minimal for very simple, latency-sensitive requests;
• low for straightforward instruction following and high-throughput work;
• medium for a balanced quality/latency tradeoff;
• high for difficult coding, multimodal reasoning, mathematics, and tool-heavy workflows.

Google retains the older thinking_budget parameter for compatibility but recommends thinking_level. The two cannot be used in the same request.

The Flash Model Can Work Through Interfaces, Not Only APIs

Computer Use is relevant when the system being automated does not expose the required operation through a clean API.

An agent can inspect visual state, decide the next action, interact with controls, observe the result, and continue.

Potential workloads include:

• browser automation;
• application testing;
• repetitive enterprise workflows;
• data entry;
• cross-system operations;
• UI-based research.

The feature remains in preview.

Production systems should evaluate more than successful demos.

Measure incorrect clicks, navigation loops, destructive actions, recovery after unexpected UI state, confirmation requirements, and successful termination.

The First Stable Successors Changed the Operating Profile

Google’s Gemini 3.5 Flash migration guide explicitly documents the move from:

gemini-3-flash-preview → gemini-3.5-flash

The guide calls out several changes:

• default thinking moved from high to medium;
• thinking_level is preferred over thinking_budget;
• older sampling overrides such as temperature, top_p, and top_k should be removed from the migration path;
• function responses need correct IDs and names;
• prompts should be re-tested for quality, speed, and cost.

Google’s current deprecation table now recommends gemini-3.6-flash as the replacement for the preview endpoint.

That means new migrations should evaluate the current stable Flash family rather than assuming 3.5 is still the best destination.

The economics also differ.

Gemini 3 Flash Preview costs $0.50 input / $3 output for standard text, image, and video usage.

Later stable Flash models can have different prices and reasoning behavior, so migration should compare complete tasks rather than only API compatibility.