Google announced Gemini 3 on November 18, 2025, calling it their most intelligent model to date. The Pro variant brings substantial improvements over Gemini 2.5, with a particular focus on multimodal understanding and agentic capabilities.

Benchmark Performance

The numbers tell a compelling story:

  • 50%+ improvement over Gemini 2.5 Pro in solved benchmark tasks
  • 81% on MMMU-Pro (multimodal understanding)
  • 87.6% on Video-MMMU
  • 72.1% on SimpleQA Verified (factual accuracy)

Google claims Gemini 3 is now the best model in the world for multimodal understanding - processing text, images, audio, and video together with improved reasoning.

1 Million Token Context Window

Gemini 3 Pro can handle up to 1 million tokens of context. This enables working with:

  • Entire code repositories
  • Long documents and PDFs
  • Hours of video content
  • Complex multi-source datasets

For developers, this means you can feed the model an entire codebase and ask questions about architecture, dependencies, or potential issues.

Generative Interfaces

One of the more interesting features is what Google calls “generative interfaces.” Rather than always responding with text, Gemini 3 can generate interactive, website-like interfaces within the app when appropriate.

Ask for travel recommendations and it might create a visual layout with modules, images, and follow-up prompts - adapting its output format to best fit the request.

New Developer Controls

Gemini 3 Pro introduces fine-grained controls for developers:

  • thinking_level: Set to low or high to balance reasoning depth against latency and cost
  • media_resolution: Control vision processing quality (low, medium, high) to manage token usage
  • Multimodal function responses: Functions can now return images and PDFs, not just text
  • Streaming function calling: Get partial function call arguments in real-time

These parameters let you tune the model’s behavior for your specific use case and cost constraints.

Agentic Capabilities

Gemini Agent is an experimental feature that handles multi-step tasks directly. It can connect to Google apps to:

  • Manage your Calendar
  • Add reminders
  • Organize your inbox

The model also brings improved tool use and agentic coding capabilities, making it more reliable for automated workflows.

Image Generation

Gemini 3 Pro Image delivers high-fidelity images with accurate text rendering. Every generated or edited image includes SynthID digital watermarks to indicate AI origin - a responsible approach to generated content.

Availability

Access Gemini 3 Pro through:

  • Gemini app (select “Thinking” from the model selector)
  • Google AI Studio
  • Vertex AI

The Competition Heats Up

With Gemini 3 Pro and Claude Opus 4.5 releasing within a week of each other, developers now have two strong options for advanced AI capabilities. Gemini leads in multimodal understanding while Claude edges ahead in coding benchmarks. The best choice depends on your primary use case.