New Gemini API updates for Gemini 3

Gemini 3, our most intelligent model, is available to developers through the Gemini API. To support its cutting-edge logic, autonomous coding and multimodal understanding, and powerful agentic capabilities, we have introduced several updates to the Gemini API. These changes are designed to give you more control over how the model reasons, how it processes media, and how it interacts with the outside world.

Here’s what’s new in the Gemini API for Gemini 3

• Simplified parameters for thinking control: Starting with Gemini 3, we are introducing a new parameter called thinking_level Controlling the maximum depth of the model’s thinking process before generating a response. Gemini 3 treats these levels as relative guidelines for reasoning rather than strict token guarantees. The thinking_level parameter allows you to adjust the depth of the internal logic of the model. You can set it to “High” for complex tasks that require optimal thinking (e.g. strategic business analysis, or scanning code for vulnerabilities). You can set it to “Low” for latency and cost-sensitive applications such as structured data extraction, summarization, etc. read more Here

• Granular control over multimodal vision processing: The media_resolution parameter lets you configure how many tokens are used for image, video, and document inputs, allowing you to balance visual fidelity with token usage. The resolution can be set for each individual media part or globally using media_resolution_low, media_resolution_medium, or media_resolution_high. If not specified, uses model Optimal defaults based on media typeHigher resolution improves the model’s ability to read fine text or identify small details, but increases token usage and latency,

• Signatures considered to improve function calling and image generation performance: Starting with Gemini 3, we are implementing the return of “Idea signature.“. These are encrypted representations of the model’s internal thought process. By sending these signatures back to the model in subsequent API calls, you ensure that Gemini 3 maintains its chain of logic throughout the interaction. This is important for complex, multi-step agentic workflows where preserving the “why” behind a decision is as important as the decision itself. If you Use official SDK And standard chat history, thought signatures are automatically managed for you. Validations on API work as follows
  • Function calling has strict validation oncurrent twist“. Missing signatures will result in a 400 error. Please read this to understand how signatures appear for different function calling scenarios. Here
  • For text/chat generation, validation is not strictly enforced, but omitting signatures will degrade the model’s reasoning and response quality.
  • Considerations in image creation/editing are strict validation for all model parts including signatures. A 400 error will occur if the signature is missing.

• Grounding and URL references with structured output: You can now combine tools hosted exclusively by Gemini Grounding with Google search and URL referencing with structured output. It’s especially powerful for building agents who need to fetch live information from the web or specific webpages and extract that data into a precise JSON format for downstream tasks.

• Updates to Grounding with Google Search Pricing: To better support dynamic agentic workflows, we are changing our pricing model from a flat rate (US$35/1k prompt) to a more granular, usage-based rate of US$14 per 1,000 search queries.

Best practices for using Gemini 3 Pro via our API

We have seen widespread enthusiasm for Gemini 3 Pro, particularly for vibe coding, zero-shot generation, mathematical problem solving, complex multimodal understanding challenges, and many other use cases. To get the best results by pushing the boundaries of Gemini 3. More Details Here,

• temperature: We strongly recommend keeping the temperature parameter at its default value of 1.0
• Consistency and defined parameters: Maintain a consistent structure throughout your signals (for example, standardized XML tags) and clearly define ambiguous terms.
• Output Verbosity: By default, the Gemini 3 is less verbose and prefers to provide direct, efficient answers. If you need a more interactive or “talkative” response, you should explicitly ask for it.
• Multimodal Compatibility: Text, images, audio or video should all be treated as the same category of input. Specific modalities should be clearly referenced in the instructions to ensure that models are synthesized across them rather than analyzed in isolation.
• Obstacle Placement: Place the behavioral constraints and role definitions in the system instruction or at the top of the prompt to ensure that they provide the basis for the model’s reasoning process.
• Long reference structure: When working with larger references (books, codebases, long videos), place your specific instructions here Ending of the prompt (after the data reference).

Gemini 3 Pro is our most advanced model agentic codingTo help developers take advantage of its best capabilities, we have worked closely with our research team system instructions template For a model that improved performance on several agentic benchmarks.

To get started building with these new features, check out gemini 3 documentation read more developer guide For technical implementation details.