# Google Unveils Gemini 3: What Changes for Developers
Table of Contents
People who build software or set up automations find themselves, every few months, facing a new AI model billed as the most capable ever. On November 18, 2025, Google announced Gemini 3, the new generation of its model family. It’s worth looking past the headline: what actually improves, for which uses, and with what caveats. This article tries to sum up the announcement soberly, separating stated facts from marketing promises.
What Gemini 3 is
Let’s start with the terms. Gemini is Google’s family of AI models, the one that powers the Gemini app, some Search features, and various tools for people who code. Gemini 3 is the new generation, taking over from the previous one (Gemini 2.5) as the flagship.
At the center of the launch is Gemini 3 Pro, available right away in preview. It’s a multimodal model: that means it can understand text, images, video, audio, and code within the same request, without handling them separately. It has a context window of about one million tokens. A token is a small chunk of text, roughly a syllable or a short word; the context window is how much material the model can keep in mind within a single conversation. One million tokens is a lot: it lets you hand over entire projects or long documents without splitting them by hand. The reply, on the other hand, can run up to about 65,000 tokens.
Alongside the main model, Google announced Gemini 3 Deep Think, a mode built for the hardest reasoning problems. At the time of the announcement it wasn’t yet broadly available: it’s aimed initially at Ultra plan subscribers. That’s worth noting, because an announced feature isn’t a feature you can already use.
The stated improvements
Google says Gemini 3 Pro beats the previous model on the main benchmarks, the standardized tests used to measure models. A few numbers are worth reporting, but with due caution: these are figures stated by the party selling the product, each one measures only a slice of the real work, and none tells the whole story.
On reasoning, Google cites 91.9% on GPQA Diamond, a test of graduate-level science questions, and 37.5% on “Humanity’s Last Exam,” a hard exam spanning many subjects.
On coding, the numbers are about concrete tasks. On SWE-bench Verified, which measures the ability to fix real bugs pulled from GitHub projects, the model claims 76.2%. On Terminal-Bench 2.0, which rates completing tasks inside a terminal, 54.2%.
Then there are the human-vote leaderboards, where people compare answers from different models without knowing which is which. Google reports topping LMArena (1501 Elo points) and WebDev Arena (1487), the latter focused on building small web apps. These are useful signals to get a sense of things, not definitive proof.
Why it matters for agents and automation
An agent is a program that, beyond answering, takes concrete actions on your behalf: it reads files, runs commands, uses external tools often connected through open standards like MCP. Anyone evaluating a model to build automation looks at exactly this, not just the quality of the text it produces.
That’s why benchmarks like Terminal-Bench or SWE-bench weigh more than others: they show how the model does when it has to move among files, commands, and tools, rather than just chatting. Along with the model, Google introduced Antigravity, a development platform built to let agents work with direct access to the editor, the terminal, and the browser. Gemini 3 also lands inside Gemini CLI, the existing command-line assistant.
For people designing systems, one more technical detail matters: the API introduces a thinking level parameter, which sets how deeply the model should “reason” before answering. It’s a practical lever: reasoning more can improve quality on complex problems, but it costs more and takes longer. Being able to tune it helps balance spend, speed, and accuracy for each task.
How to try it and what it costs
For developers, Gemini 3 Pro is reachable through the Gemini API in two places: Google AI Studio, with free access subject to usage limits, and Vertex AI, the cloud platform aimed at businesses. The model is also built into third-party tools already popular among developers, such as Cursor, GitHub, JetBrains, and Replit. For the general public, it arrived in the Gemini app and in some Search features.
On costs, Google listed a price of 12 per million output tokens, for requests up to 200,000 tokens; past that threshold the rate changes. These figures help estimate the cost of heavy use, but should be checked against the documentation at the time of adoption, since model pricing changes often.
The caveats
Before moving over to the new model, a few points are worth keeping in mind, without hype.
It’s a preview. Google describes it as a preview release: behavior, features, and prices can change. It’s not the thing to hang critical processes on without a ready fallback.
The numbers are one-sided. The benchmarks are stated by the party selling the product and measure standardized scenarios. The only reliable judge is trying it on your own real use cases.
Data goes to the cloud. As with the earlier models, the computation runs on Google’s servers: whatever you send the model leaves your own computer. For confidential content, that’s something to weigh carefully, also in light of what GDPR requires on the infrastructure side when personal data is involved.
Not everything is available at once. Deep Think and some features weren’t accessible to everyone at launch. It’s best to check what you can actually use, and on which plan, before relying on a specific capability.
In short
Gemini 3 is the new generation of Google’s models, with Gemini 3 Pro available in preview and a Deep Think mode announced for the hardest reasoning tasks. Google claims improvements in reasoning, coding, and multimodal understanding, with interesting but one-sided numbers. For those building agents and automations, the most concrete news is the better performance on practical tasks, the Antigravity platform, and the control over reasoning depth through the API. It remains a preview tool, running in the cloud, to be measured on your own cases: worth trying, without taking the press releases at face value.
