Google Gemini 3 Pro and Antigravity IDE: What Changed

On November 18, 2025, Google released Gemini 3 Pro and Google Antigravity. Gemini 3 Pro is their latest language model. Antigravity is a development environment built around autonomous AI agents. Both are available now.

Gemini 3 Pro benchmark results

Gemini 3 Pro is the first model in the Gemini 3 series. The model has a 1 million token input context window and supports up to 64k tokens of output.

Benchmark scores

Here's how Gemini 3 Pro performed on standard benchmarks:

LMArena Leaderboard: 1501 Elo (currently first place)
Humanity's Last Exam: 37.5% without tools
GPQA Diamond (scientific knowledge): 91.9%
MathArena Apex: 23.4%
AIME 2025: 100% with code execution
ARC-AGI-2 (visual reasoning): 31.1%
MMMU-Pro (image understanding): 81%
Video-MMMU: 87.6%
SimpleQA Verified (factual accuracy): 72.1%

The model handles text, images, video, audio, and code in the same context. Its spatial understanding works for pointing, trajectory prediction, and task progression, which matters for robotics and autonomous vehicle applications.

Coding benchmarks

Coding-specific results:

Terminal-Bench 2.0 (terminal operations): 54.2%
SWE-bench Verified (coding agents): 76.2%
WebDev Arena: 1487 Elo
LiveCodeBench Pro: 2439 Elo rating

The model can generate applications from natural language descriptions and handle multi-file refactoring. It maintains context across debugging sessions better than Gemini 2.5 Pro.

Deep Think mode

Gemini 3 Deep Think is a variant that allocates more compute to reasoning before responding. It scores higher on most benchmarks:

Humanity's Last Exam: 41.0% (vs 37.5% for standard)
GPQA Diamond: 93.8% (vs 91.9%)
ARC-AGI-2 with tools: 45.1% (vs 31.1%)

Deep Think will be available to Google AI Ultra subscribers after additional safety testing.

Google Antigravity IDE

Google Antigravity is a development environment designed around autonomous agents. It's free during public preview and runs on MacOS, Windows, and Linux.

Two-part interface

Antigravity has two main views:

Editor view: A standard IDE with tab completion, inline commands, and an agent in the side panel. This works like existing AI-powered editors.

Manager view: A control panel for running multiple agents across different workspaces. You can start agents for background tasks, monitor their progress through an inbox, and switch between workspaces. The agents run asynchronously rather than blocking your work.

How agents work

Agents in Antigravity operate across the editor, terminal, and browser at the same time. They can write code, start servers, and test the results in a browser without waiting for input.

The agent outputs artifacts as it works: task lists, implementation plans, screenshots, and browser recordings. These make it easier to check what the agent did without reading through individual API calls.

You can comment on any artifact, similar to commenting in Google Docs. The agent reads these comments and adjusts its work without stopping. Comments work on text artifacts and screenshots.

Agents maintain a knowledge base. They save useful patterns and code snippets as they work, then retrieve this information for later tasks. The system tracks both explicit knowledge (like snippets) and implicit patterns (like successful task sequences).

Browser control

Antigravity uses Gemini 2.5 Computer Use for browser automation. Agents can open browsers, navigate pages, and verify that applications work correctly. This happens automatically during testing workflows.

Available models

Antigravity supports multiple models:

Gemini 3 Pro (Google)
Claude Sonnet 4.5 (Anthropic)
OpenAI models

Rate limits reset every five hours and are based on the amount of work done by the agent. Google states that most users won't hit these limits during normal usage.

API changes

Thinking level parameter

The thinking_level parameter controls how much compute the model uses before responding:

low: Faster responses, lower cost, suitable for simple tasks
high (default): More reasoning time, better for complex problems
medium: Coming later

You cannot use thinking_level and the older thinking_budget parameter together.

Media resolution

The media_resolution parameter sets token allocation for images and videos:

| Content | Recommended | Tokens | Notes | | --------------- | ----------- | ------------- | ------------------------------- | | Images | high | 1120 | Default for most image tasks | | PDFs | medium | 560 | Quality doesn't improve at high | | Video (general) | low/medium | 70 per frame | Same token count for both | | Video (OCR) | high | 280 per frame | For reading text in videos |

Thought signatures

Gemini 3 Pro returns encrypted "thought signatures" in its responses. You must send these back in follow-up requests to maintain reasoning context. The official Python, Node, and Java SDKs handle this automatically.

If you're using the API directly, include the thoughtSignature field from the previous response in your next request. Missing signatures cause 400 errors for function calls and image generation, and degraded quality for text generation.

Tool integration

Gemini 3 Pro can use structured outputs with built-in tools:

Google Search (real-time information)
URL Context (fetch web pages)
Code Execution (run Python)
File Search (query uploaded documents)

This works with function calling and JSON schema outputs.

Pricing and access

Pricing for gemini-3-pro-preview:

| Context size | Input | Output | | ------------- | ----- | ------ | | < 200k tokens | $2/M | $12/M | | > 200k tokens | $4/M | $18/M |

Available through:

Google AI Studio (free tier with rate limits)
Vertex AI
Gemini CLI
Google Antigravity
Cursor, GitHub Copilot, JetBrains, Manus, Replit, Cline

What people are building

Google demonstrated several single-prompt applications: a 3D spaceship game, procedural fractal generators, and voxel art tools. These run entirely in the browser without additional setup.

Cline, an autonomous coding assistant, reports that Gemini 3 Pro handles multi-file refactoring better than other models. It maintains context through longer sessions and solves problems that caused other models to fail.

OpusClip, a video processing service, reports 32% faster processing compared to their previous implementation. They cite better instruction following and less hallucination in long-context scenarios.

Emergent, a fullstack development platform, uses Gemini 3 Pro's multi-step tool calling for UI generation. They report better prompt adherence compared to previous models.

Safety testing

Google tested Gemini 3 Pro against their Frontier Safety Framework and provided early access to the UK AISI. Third-party assessments came from Apollo, Vaultis, and Dreadnode.

The model shows less sycophancy (agreeing with users regardless of correctness) and better resistance to prompt injection attacks compared to Gemini 2.5 Pro.

Knowledge cutoff is January 2025. Use Google Search grounding for more recent information.

Migrating from Gemini 2.5 Pro

Key changes:

Temperature: Keep it at 1.0. Lower values can cause looping or worse performance on complex tasks. This is different from earlier models.

Prompting: Remove chain-of-thought prompting if you were using it. Set thinking_level: "high" and simplify your prompts instead.

PDF handling: Default OCR resolution changed. Test your document parsing workflows with the new defaults or adjust media_resolution.

Token usage: PDF token consumption may increase, video token usage may decrease compared to Gemini 2.5 Pro.

Features: Context Caching requires minimum 2,048 tokens. Batch API is supported. Image segmentation is not supported (use Gemini 2.5 Flash or Robotics-ER 1.5 instead).

Integration with existing tools

Google AI Studio

Google AI Studio's Build mode connects models and APIs automatically. The interface includes code annotations for faster iteration. There's an "I'm feeling lucky" button that generates app concepts and implementation at the same time.

Android Studio and CLI

Android Studio includes Gemini 3 Pro for code generation. The Gemini CLI tool lets the model propose shell commands for filesystem navigation and build automation.

Third-party tools

These platforms support Gemini 3 Pro at launch:

Cline (autonomous coding)
OpusClip (video processing)
Emergent (fullstack development)
Cursor, JetBrains, Manus, Replit

What's next

Gemini 3 Deep Think will be available to Google AI Ultra subscribers after more safety testing. Google plans to release additional Gemini 3 models but hasn't announced a timeline.

Consumer rollout:

Gemini app: Available to all users
Google Search AI Mode: Pro and Ultra subscribers
Gemini Agent: Handles multi-step tasks like email organization (Ultra subscribers)

Summary

Gemini 3 Pro beats Gemini 2.5 Pro on standard benchmarks, particularly for coding and multimodal tasks. The 1M token context window and native support for multiple input types (text, images, video, audio, code) make it suitable for complex applications.

Google Antigravity introduces a two-view interface: a standard editor for direct work and a manager for asynchronous agent orchestration. Agents can work across the editor, terminal, and browser simultaneously. The system outputs artifacts (plans, screenshots, recordings) rather than raw tool calls, and accepts feedback through inline comments.

Both are available now. Gemini 3 Pro has a free tier in Google AI Studio with rate limits. Antigravity is free during public preview with five-hour rate limit resets.

The main tradeoffs: Gemini 3 Pro requires returning thought signatures in API calls and performs worse with temperature below 1.0. Antigravity's agent-first approach means you need to adapt to asynchronous workflows rather than real-time collaboration.