Google Rolls Out New Gemini API Features to Support Gemini 3

Google has introduced several updates to the Gemini API to better support its new Gemini 3 model. The changes focus on easier control over how the model thinks, improved options for handling images and video, and the return of “thought signatures” to boost function calling and image generation.

The company detailed the updates in a November 25 blog post. According to Google, the improvements give developers more control over how Gemini 3 reasons, processes multimodal content, and interacts with external tools. Gemini 3 itself launched on November 18.

One major update is a simplified way to control the model’s thinking process. A new parameter called thinking_level sets how deep the model’s internal reasoning can go before producing an answer. A high setting is meant for complex work, like strategic or analytical tasks, while a low setting helps reduce cost and latency for simpler requests.

The API also adds more precise controls for vision tasks. A new media_resolution option lets developers decide how many tokens the model should spend on images, videos, or documents. It can be set to low, medium or high. Higher settings improve the model’s ability to read small text or pick out fine details, although they use more tokens.

Another key change is the return of thought signatures. These are encrypted snapshots of the model’s internal reasoning. Developers can pass them back into later API calls so the model remembers how and why it reached past decisions. This helps in long, multi-step workflows where consistent reasoning is important.

Google also now allows developers to use structured outputs together with Gemini-hosted tools such as Grounding with Google Search and URL Context. This combination is useful for building agents that pull live information from the web and convert it into clean JSON for other tasks.

Along with the feature updates, Google adjusted the pricing for Grounding with Google Search. Instead of a flat $35 per 1,000 prompts, the cost is now usage-based at $14 per 1,000 search queries.