Google updates Gemini API for Gemini 3

Google has updated its Gemini API in support of its recently introduced Gemini 3 AI model. The Gemini API updates, which include simpler controls over thinking, more granular control over multimodal vision processing, and ‘thought signatures’ to improve function calling and image generation, are designed to support Gemini 3’s reasoning, autonomous coding, multimodal understanding, and agentic capabilities.

The Gemini API improvements were detailed in a November 25 blog post. They are designed to give users more control over how the model reasons and processes media, and how the model interacts with the “outside” world, Google said. Gemini 3 was rolled out November 18.

Among the improvements in the Gemini API are simplified parameters for thinking control, via a parameter called thinking_level. This parameter controls the maximum depth of the model’s internal reasoning before a response is produced. The thinking_level parameter can be set to high for complex tasks such as strategic business analysis, and low for latency and cost-sensitive applications.

The Gemini API also now provides more granular control over multimodal vision processing, with a media_resolution parameter for configuring how many tokens are used for image, video, and document inputs. Developers can balance visual fidelity with token usage. Resolution can be set using media_resolution_low, media_resolution_medium, or media_resolution_high. Higher resolution boosts the model’s ability to read fine text or identify small details, Google said.

Starting with Gemini 3, Gemini API also brings back thought signatures to improve function calling and image generation. Thought signatures are encrypted representations of the model’s internal thought process. By passing these signatures back to the model in subsequent API calls, developers can ensure that Gemini 3 maintains its chain of reasoning across a conversation. This is important for complex, multi-step agentic workflows where preserving the “why” behind a decision is as important as the decision itself, Google said.

Additionally, developers now can combine structured outputs with Gemini-hosted tools, specifically Grounding with Google Search and URL Context. Combining structured outputs is especially powerful for building agents that must fetch live information from the web or specific web pages and extract the data into a JSON format for downstream tasks, Google said, noting that it has updated the pricing of Grounding with Google Search to better support agentic workflows. The pricing model changes from a flat rate of $35 per 1K prompts to a usage-based rate of $14 per 1,000 search queries.