Google has announced the release of Gemini 3 Pro, claiming it to be their most intelligent foundation model, demonstrating state-of-the-art performance across major reasoning (e.g., Humanity's Last Exam) and multimodal agentic coding benchmarks. The model features a native 1 million-token context window and introduces granular API controls over internal reasoning depth (thinking_level) and multimodal resolution to manage latency and cost. Concurrently, Google is launching Google Antigravity, a new agent-first development platform designed to leverage Gemini 3 Pro's improved agentic capabilities for autonomous, complex software development and "vibe coding." A more potent, safety-delayed variant, Gemini 3 Deep Think, promises further advancements in solving complex problems.
Summary of the Gemini 3 Announcement:
Model Release and Tiers: Gemini 3 Pro is the initial release in preview, accessible today via the Gemini app, Search, AI Studio, and Vertex AI. It offers a 1 Million (1M) input token and 64k output token context window, with a knowledge cutoff of January 2025.
Performance Claims (State-of-the-Art): Google claims Gemini 3 Pro sets new highs, including topping the LMArena Leaderboard (1501 Elo). Notable benchmark scores (no tools):
Academic Reasoning (Humanity's Last Exam): $37.5%$
Long-Horizon Agentic Tasks (Vending-Bench 2 Net Worth): $$5,478.16$
Deep Think Mode: Gemini 3 Deep Think is an enhanced reasoning mode ($41.0%$ on Humanity's Last Exam) intended for the most complex problems. It will be released to Google AI Ultra subscribers later, pending additional safety evaluations.
New API Controls for Efficiency: Developers now have explicit control over model behavior:
Thinking Level: The thinking_level parameter (low/high) allows developers to constrain the model's internal reasoning depth, optimizing latency and cost when complex reasoning is not required.
Media Resolution: Granular control over multimodal vision processing via media_resolution (low/medium/high) is introduced to balance visual fidelity against token consumption and latency for images and video frames.
Pricing: Gemini 3 Pro is priced at $$2/\text{million input tokens}$ and $$12/\text{million output tokens}$ for prompts under $200k$ tokens, as far as I can tell from the provided developer documentation.
Agentic Development Platform (Google Antigravity): Google launched Antigravity, a new agent-first IDE that uses Gemini 3 Pro agents to autonomously plan and execute complex, end-to-end software tasks across the editor, terminal, and browser. This is presented as elevating the developer experience to a task-oriented level.
Vibe Coding and Multimodality: The model is optimized for "vibe coding" (translating high-level natural language into interactive apps) and demonstrates improved multimodal understanding, including document parsing and screen understanding (ScreenSpot-Pro score of $72.7%$).
API Requirements: The API enforces the use of "Thought signatures"—encrypted representations of the model's internal reasoning process—which must be returned in subsequent calls to maintain optimal performance, especially for strict function calling workflows. Failing to return them for critical tasks results in a $400$ error, which sounds like unnecessary complexity but is deemed critical for complex multi-turn reasoning.
Safety and Responsibility: Google asserts Gemini 3 is their most secure model yet, showing reduced sycophancy and increased resistance to prompt injections, having undergone extensive in-house and external safety evaluations.