Gemini 3 Pro Preview API

Google's latest agentic model featuring Thinking Mode, Thought Signatures for stateful reasoning, and native multimodal understanding.

Playground Not Available

This feature is currently only available for selected image and video generation models.

Gemini 3 Pro Preview API - Thinking Mode Enabled

Deploy gemini-3-pro-preview with configurable reasoning depth. Utilize Thought Signatures for complex agentic workflows and handle 1M+ tokens via EvoLink.

Capabilities of Gemini 3 Pro Preview API

Deep Thinking Mode

Configure 'Thinking Levels' (Low/High) to balance latency with reasoning depth.

Explore Logic

Agentic & Thought Signatures

Pass encrypted Thought Signatures to maintain reasoning context across multi-step turns.

Build Agents

1M Multimodal Context

Process vast datasets including video, audio, and code repositories in a single pass.

Test Capacity

Why Integrate Gemini 3 Pro via EvoLink

Leverage the new gemini-3-pro-preview with enterprise-grade caching and routing. We optimize the tiered pricing structure automatically.

Thinking Level Control

Dynamically switch between fast instruction following and deep 'High' thinking mode via API parameters.

Reliable Orchestration

Use Thought Signatures to ensure your agents remember their internal logic steps during function execution.

Cost-Efficient Scaling

Access the $2/$12 rate for standard prompts (<200k) and utilize EvoLink caching to minimize repeat input costs.

How to Use Gemini 3 Pro Preview API

Set your Thinking Level, define tools, and route via EvoLink.

Step 1 - Configure Model

Select 'gemini-3-pro-preview' and set `thinking_level` to 'High' for complex tasks or 'Low' for speed.

Step 2 - Integrate Signatures

Ensure your API loop captures and returns `thought_signature` tokens to maintain context state.

Step 3 - Optimize & Run

Send requests via EvoLink. We handle the 200k tier routing and cache static context to save ~90% on inputs.

Open Developer Playground

Technical Specs

Advanced features of the Gemini 3 Pro Preview API

Context

1M Token Window

Massive context for analyzing books, legal discovery, or hours of video.

Reasoning

Thinking Mode

New API parameter to control the depth of the model's internal chain-of-thought.

Cost

Tiered Pricing

Pay $2/$12 per 1M tokens for inputs under 200k; $4/$18 for extended context.

Architecture

Thought Signatures

Encrypted tokens that preserve reasoning state across API calls for agents.

Multimodal

Native Multimodal

First-class support for PDF, Audio, Video, and Image inputs without OCR middleware.

Performance

Prompt Caching

Reduce latency and cost for repetitive large contexts via EvoLink.

Gemini 3 Pro vs Competitors

Choosing the right reasoning model

Model	Duration	Resolution	Price	Strength
Gemini 3 Pro Preview	N/A	Thinking Mode	$2/$12 (<200k tokens)	Adjustable reasoning depth, Thought Signatures, 1M context.
Gemini 2.5 Pro	N/A	Standard CoT	~$1.25/$10	Solid baseline performance if 'Deep Thinking' is not required.
GPT-5 / Opus	N/A	Blackbox	Typically Higher	Strong reasoning but often lacks the granular control of Thinking Levels.

Gemini 3 Pro API FAQs

Everything you need to know about the product and billing.

Pricing is tiered based on context size. For requests under 200k tokens, it is ~$2.00 (input) / ~$12.00 (output) per 1M tokens. For requests over 200k tokens, rates increase to ~$4.00 / ~$18.00.

You can set a `thinking_level` parameter. 'High' allocates more compute for complex reasoning/math (slower but more accurate), while 'Low' is optimized for standard latency.

They are encrypted tokens returned by the model representing its internal reasoning. You must pass these back in subsequent API calls to ensure the model maintains its 'train of thought'.

No, the `gemini-3-pro-preview` is a text/multimodal-input model. For image generation, use the separate Gemini 3 Pro Image Preview variants.

Yes, benchmarks show significant improvements in SWE-Bench (coding) and math. The addition of Thinking Mode allows it to handle complex logic that 2.5 Pro struggles with.