Google has launched Gemini 3, aggressively marketed as their "most intelligent model" and a significant step toward AGI. The core offering is Gemini 3 Pro, which purportedly achieves state-of-the-art results across major benchmarks, leveraging advanced reasoning, 1 million token context, and world-leading multimodal capabilities. A distinct, more powerful reasoning mode, Gemini 3 Deep Think, is also announced but restricted for early safety testing. Crucially, the launch includes a new agentic development platform, Google Antigravity, and introduces strict API controls (like Thought Signatures and Thinking Levels) necessary for managing the model's heightened internal complexity and mandatory sequential reasoning during tool use.
Gemini 3: Claims, Capabilities, and Deployment
0:00 Introduction of Gemini 3 Pro: Google's latest foundation model, released in preview, focuses on state-of-the-art reasoning, multimodal understanding, and advanced agentic/coding capabilities.
0:35 Immediate Deployment: Gemini 3 is integrated into Google Search (AI Mode, shipping on day one), the Gemini app, AI Studio, Vertex AI, and the new Google Antigravity platform.
1:57 State-of-the-Art Benchmark Claims: Google claims significant performance leaps, topping the LMArena Leaderboard (1501 Elo) and achieving high scores in specialized metrics, including PhD-level reasoning on Humanity’s Last Exam (37.5% without tools) and multimodal understanding (81.0% on MMMU-Pro). These are, as far as I can tell, self-reported metrics.
2:45 Responses and Nuance: Google asserts that Gemini 3 Pro delivers "smart, concise and direct" responses that prioritize "genuine insight" over "cliché and flattery." This seems like marketing bravado until demonstrated consistently.
3:05 Gemini 3 Deep Think Mode: An enhanced reasoning mode, demonstrating further benchmark improvements (e.g., 41.0% on Humanity's Last Exam). This mode is being held back for additional safety evaluation before being made available to Google AI Ultra subscribers soon.
3:56 Core Utility (Learn, Build, Plan): The model is designed for universal utility, including decoding handwritten recipes, generating interactive visualizations from academic papers, enhancing web UI generation ("vibe coding"), and managing complex, multi-step planning tasks (long-horizon agentic tasks demonstrated on Vending-Bench 2).
4:54 Agentic Coding Prowess: Gemini 3 Pro claims superior agentic and coding performance, topping the WebDev Arena (1487 Elo) and scoring 76.2% on SWE-bench Verified.
5:12 New Agentic Development Platform (Google Antigravity): This new IDE focuses on an agent-first experience, allowing developers to collaborate with autonomous agents that plan, execute, and validate complex software tasks across the editor, terminal, and browser. It is available as a public preview at no charge for major OSs.
9:30 API Pricing and Limits: Gemini 3 Pro (preview) costs $2/$12 per million input/output tokens (for contexts under 200k tokens). It supports a massive 1 million token input context and 64k output tokens. The knowledge cutoff is January 2025.
11:58 Mandatory Thinking Control: The new thinking_level parameter controls the model's internal reasoning depth (default is "high"). Disabling thinking is not possible for Gemini 3 Pro. Crucially, API pricing is based on the full internal thought tokens generated, not just the summarized output.
12:15 Granular Multimodal Resolution: The new media_resolution parameter allows explicit control over the token count allocated for images, videos, and PDFs (low, medium, high), providing developers necessary levers to manage latency and cost associated with multimodal processing.
13:17 Strict Thought Signatures (Critical Detail): Gemini 3 introduces strict validation for thought_signatures—encrypted tokens representing the model's internal reasoning context. For multi-step function calling workflows, returning these signatures exactly as received in the conversation history is mandatory (omission results in a 400 error), indicating that sequential reasoning is now non-optional for tool use. I don't care about the specific JSON examples, but the necessity of maintaining internal state via signatures is a significant change for developers.
15:00 Migration Guidance: Developers migrating from Gemini 2.5 are advised to use simplified prompts and the new thinking_level instead of complex prompt engineering, and to remove explicit temperature settings, as the default (1.0) is now optimized for the model's reasoning capabilities. Image segmentation is no longer supported in 3 Pro.