Credit: Google

Google has released a significant set of Gemini API enhancements designed specifically for Gemini 3, the company’s most advanced model for reasoning, coding, multimodal understanding, and agent-driven automation. These updates give developers deeper control over how the model thinks, processes media, and maintains long-running chains of reasoning.

Smarter Reasoning Control with thinking_level

Developers can now guide the depth of Gemini 3’s internal reasoning through a new thinking_level parameter.

Use cases:

• high – complex analysis, vulnerability detection, multi-step planning

• medium – balanced reasoning

• low – fast tasks like extraction or summarization

Unlike strict token-based controls, Gemini treats these levels as relative reasoning guidelines, giving more predictable behavior across varied workloads.

Granular Multimodal Resolution Settings

The new media_resolution parameter gives precise control over how many tokens Gemini allocates to image, video, and document inputs.

Available options:

• media_resolution_low

• media_resolution_medium

• media_resolution_high

Higher resolution increases fidelity for fine text or small details but uses more tokens and increases latency. Resolution can be set globally or per media item.

Thought Signatures: New Requirement for Multi-Step Intelligence

Gemini 3 introduces Encrypted Thought Signatures, required for preserving coherent reasoning across multi-turn agent workflows.

What they do:

• Encode the model’s internal reasoning

• Must be returned on the next API call for continuity

• Enable more consistent function calling, tool use, and image generation

Validation rules:

• Function Calling – strict; missing signature = 400 error

• Image generation/editing – strict; signature required for every part

• Chat/Text generation – flexible; missing signature reduces quality

SDK users get this automatically; raw API users must pass signatures manually.

Grounding + URL Context + Structured Outputs

Developers can now combine:

• Google Search Grounding

• URL context ingestion

• Structured JSON outputs

This enables agentic systems that can retrieve live web data, extract structured information, and feed it into downstream workflows with high precision.

Updated Grounding with Google Search Pricing

Google is moving from a flat rate to usage-based pricing:

$14 per 1,000 search queries

(previously $35 per 1,000 prompts)

This better supports agent-heavy applications that trigger frequent search actions.

Gemini 3 Pro excels in zero-shot generation, coding agents, math, long-context reasoning, and multimodal synthesis.

Developers can start building with the new features today by visiting the Gemini 3 documentation and the Developer Guide for implementation details.