Introduction to the New Gemini 3 API
The evolution of artificial intelligence takes a new leap forward with the introduction of the Gemini 3 API. Google DeepMind has released a series of significant updates designed to give developers unprecedented control over the model's reasoning capabilities, autonomous coding, and multimodal understanding. These new features are not just additions, but fundamental tools for those building complex and flexible AI agents.
The Gemini 3 API is enriched with parameters to manage the depth of the model's thinking and security mechanisms to preserve logical consistency across conversations. In this article, we will explore in detail how to use Thinking Levels, Thought Signatures, and new multimodal configuration options to maximize your application's performance.
Reasoning Control: Thinking Levels
One of the most relevant innovations is the introduction of the thinking_level parameter. This feature allows developers to modulate the depth of the model's internal reasoning process before a response is generated. Unlike strict token limits, these levels act as relative guidelines for the AI's cognitive behavior.
- High Level: Ideal for complex tasks requiring deep strategic analysis, such as vulnerability scanning in code or business analysis. The model dedicates more computational resources to process the optimal response.
- Low Level: Perfect for latency and cost-sensitive applications, such as structured data extraction or fast summarization, where speed is prioritized over speculative depth.
Granular Media Management
With the new Gemini 3 API, control over costs and performance also involves visual input management. The media_resolution parameter allows configuring how many tokens are allocated for processing images, videos, and PDF documents.
Developers can choose between media_resolution_low, media_resolution_medium, or media_resolution_high. This granularity is applicable globally or for individual media parts. Increasing resolution improves the model's ability to read fine text or identify small details but entails an increase in token usage and latency. If unspecified, the system applies optimal defaults based on the media type.
Thought Signatures: Preserving the Chain of Thought
To ensure the model maintains logical consistency during complex workflows, Google has introduced "Thought Signatures." These are encrypted representations of the model's internal thought process.
By returning these signatures to the model in subsequent API calls, you ensure that Gemini 3 does not lose its train of thought. This is critical for multi-step agentic workflows, where the "why" behind a decision is as important as the decision itself. Official SDKs handle this process automatically, but for those using the API directly, here are the validation rules:
- Function Calling: Requires strict validation on the current turn. Missing signatures generate a 400 error.
- Image Generation/Editing: Strict validation; missing signatures block the request.
- Chat/Text: Validation is not blocking, but omitting signatures significantly degrades reasoning quality.
- Gemini 3 API usage requires careful handling of these signatures to maintain state.
Grounding, Structured Outputs, and Pricing
Integration between Grounding tools (like Google Search) and structured outputs is now fully supported. This allows building agents that fetch live information from the web and format it immediately into precise JSON for downstream tasks.
In parallel, the pricing model for Grounding with Google Search has transitioned from a flat rate to a usage-based model: $14 per 1,000 search queries, offering greater flexibility for dynamic workflows.
Best Practices for Developers
To get the most out of the Gemini 3 API, Google recommends some essential practices:
- Temperature: Keep the default value at 1.0.
- Consistency: Use uniform structures in prompts (e.g., standardized XML tags).
- Verbosity: Gemini 3 tends to be concise. If a conversational response is needed, it must be explicitly requested.
- Long Context: Place specific instructions at the end of the prompt, after context data (books, codebases), to better anchor reasoning.
Useful Resources
To dive deeper into technical details and start implementation, check out the Gemini 3 Developer Guide and documentation.
Conclusion
Updates to the Gemini 3 API mark an important step towards more autonomous and reliable AI agents. The combination of reasoning control, efficient media management, and security via Thought Signatures offers developers a powerful toolkit to build the next generation of intelligent applications.
FAQ
What are Thinking Levels in the Gemini 3 API?
They are parameters that allow setting the model's reasoning depth to "high" for complex tasks or "low" for speed and efficiency.
What are Thought Signatures used for?
They serve to preserve the model's chain of reasoning across multiple interactions, ensuring consistency in complex agentic workflows.
How does Grounding pricing change?
Pricing transitions from a flat rate to a usage-based model of $14 per 1,000 Google search queries.
Can I control the resolution of images sent to the API?
Yes, via the media_resolution parameter, it is possible to balance visual fidelity and token consumption.
What is the recommended temperature for Gemini 3?
Google recommends keeping the temperature at the default value of 1.0 to achieve the best performance.