On November 18, 2025, Google released Gemini 3, which it called its most intelligent model, starting with Gemini 3 Pro in preview for developers and enterprises. Google said the model topped the LMArena leaderboard with a score of 1501 Elo and posted large gains in reasoning, multimodal understanding and agentic coding over Gemini 2.5 Pro. It launched alongside Google Antigravity, a new agentic development platform, and new API controls that change how developers tune reasoning depth and image processing.
Key Facts#
- Release date: November 18, 2025. Developers could build with Gemini 3 Pro in preview (
gemini-3-pro-preview) on Vertex AI, in Google AI Studio and through the Gemini CLI, and enterprises through Gemini Enterprise. - Context: a 1-million-token context window, which Google said leads the industry on long-context performance.
- Preview pricing per million tokens: $2 input and $12 output for prompts up to 200,000 tokens, and $4 and $18 above that.
- Vendor-reported results: 1501 Elo on LMArena, 37.5% on Humanity's Last Exam without tools, 91.9% on GPQA Diamond, 31.1% on ARC-AGI-2, 76.2% on SWE-bench Verified, 54.2% on Terminal-Bench 2.0 and 81% on MMMU-Pro.
- Deep Think: an enhanced reasoning mode that Google reported at 41% on Humanity's Last Exam and 45.1% on ARC-AGI-2.
- Partner results: GitHub reported 35% higher accuracy than Gemini 2.5 Pro in resolving software engineering challenges during early testing in VS Code, and JetBrains reported more than a 50% improvement in solved benchmark tasks.
- Ecosystem: Cursor, GitHub, JetBrains, Manus and Replit were among the coding platforms integrating Gemini 3 Pro at launch.
What Happened#
Google positioned Gemini 3 as a model for doing, not only answering. Its enterprise announcement emphasized agentic coding, the ability to take in entire codebases through the million-token window, better front-end and user interface generation, and more reliable planning across multistep tool use. The benchmark gains over Gemini 2.5 Pro were large by any standard: Humanity's Last Exam rose from roughly a fifth of questions answered correctly to more than a third, and SWE-bench Verified moved into the mid-seventies.
The launch was also a platform event. Google introduced Antigravity, a development environment built around agents that can work across an editor, a terminal and a browser, and made Gemini 3 available in the Gemini CLI on day one. Third-party tools followed immediately, which meant many developers first met Gemini 3 inside the coding assistant they already used rather than through the API.
Gemini 3 also changed several API conventions. Instead of the numeric thinking budget used by Gemini 2.5, Gemini 3 models use a thinking_level setting, and Google says requests cannot combine the two. A new media_resolution parameter controls how many tokens each image or video frame consumes. Google also tightened the handling of thought signatures, the encrypted tokens that preserve the model's reasoning state across function calls: if a raw request omits them in multi-turn tool use, the API returns an error. Google additionally recommends leaving temperature at its default of 1.0 for Gemini 3, because the model's reasoning is tuned for that setting.
Background#
Gemini 3 arrived about five months after Google made Gemini 2.5 Pro and Flash generally available, and in the middle of a dense release cycle. OpenAI had shipped GPT-5.1 on November 12, 2025, a week earlier, and Anthropic's Claude Sonnet 4.5 had set a high bar for coding agents in September. With Gemini 3, Google again claimed the top of the LMArena leaderboard, as it had with Gemini 2.5 Pro in the spring.
The rollout was also broader from the start. Gemini 3 Pro reached developers, enterprises and third-party coding tools on launch day, with a new agent platform to showcase it.
Why It Matters for Developers#
For .NET teams, the most practical lesson is in the API changes. If you built on Gemini 2.5 with thinking budgets, moving to Gemini 3 means switching to thinking levels, and you should drop any temperature tuning you relied on before. Keep these provider-specific settings in configuration, so a model upgrade does not require a code change.
Thought signatures deserve a careful look. Google's guidance says its Python, Node.js, Go and Java SDKs, and its OpenAI-compatible Chat Completions endpoint, handle them automatically when you send back the full conversation history. If you call Gemini from C#, through the Google.GenAI .NET SDK and its IChatClient implementation or through your own HTTP code, test multi-turn function calling explicitly before you ship. A missing signature is exactly the kind of bug that passes single-turn tests and fails in production.
Abstractions help here. With Microsoft.Extensions.AI, your application code talks to IChatClient, so you can compare Gemini 3 with other models on the same evaluation suite and switch providers without rewriting business logic. That matters because new frontier models from every major vendor followed within months.
Finally, the partner results from GitHub and JetBrains suggest that developers using Copilot or JetBrains AI tools would see Gemini 3 in their existing workflow. Our guide to AI-assisted .NET development covers how to evaluate a new model inside those tools rather than trusting vendor benchmarks.
What's Next#
Google followed with a Gemini 3 Flash preview, then Gemini 3.1 Pro in February 2026, and a Gemini 3.5 generation at Google I/O in May 2026. Gemini 3 Pro itself stayed a preview model. By September 2026, according to model-pricing notes maintained by the Langfuse project, Google's Gemini API documentation listed Gemini 3 Pro Preview as shut down, leaving Gemini 3.1 Pro as its successor.
That short lifespan is the main caution for developers. Preview models can move fast, so anything you build on them should assume a migration within months. Pin the model ID in configuration, keep an evaluation suite ready, and watch Google's model lifecycle pages rather than assuming a preview will graduate unchanged.