Google Launches Gemini 3 Pro

Google today launched Gemini 3 Pro, the latest iteration of its family of foundation models, with other Gemini 3 variants to follow. Gemini 3 is immediately available across a variety of Google products, including Google Search, the Gemini App, AI Studio, Vertex AI and Google’s various agentic development tools, including the newly launched Antigravity.

Record Benchmarks

Unsurprisingly, Google describes Gemini 3 Pro as its most intelligent model yet. According to Google, the model will top the LMArena Leaderboard with a score of 1501 points. It also scores 37.5% on Humanity’s Last Exam and 91.9% on GPQA Diamond, a benchmark that asks Ph.D.-level science questions and which OpenAI’s GPT-5.1 previously led with 87.6% accuracy.

One area the Google team stresses is Gemini 3 Pro’s ability to do well on multimodal reasoning tasks, with the model reaching record scores on many of those benchmarks as well. This also means it’ll excel in visual reasoning tasks, and another area Google highlighted is its ability to create complex animations and simulations on the fly.

Benchmarks aren’t everything, of course, and can even be a bit misleading at times, but Google argues that the new model can bring “a new level of depth and nuance to every interaction,” without resorting to cliches and the sycophantic behavior often seen in other models. It’s only in day-to-day use, though, where the differences between model families really become obvious, so it’ll take a few days for developers and consumers to get a feel for how well the model performs in real-world tasks.

Coding With Gemini 3

As for coding, Google says the new model also outperforms Gemini 2.5 Pro at coding in both agentic workflows and when handling zero-shot tasks. It’ll be available in Google’s own tools, like Gemini CLI and Android Studio’s Agent Mode, as well as through third-party tools from the likes of Cursor, GitHub, JetBrains, Manus and Cline, among others.

“Cline is using Gemini 3 to power autonomous code generation in developers’ IDEs,” said Nik Pash, Cline’s head of AI. “Gemini 3 Pro handles complex, long-horizon tasks across entire codebases, maintaining context through multifile refactors, debugging sessions and feature implementations. It uses long context far more effectively than Gemini 2.5 Pro and has solved problems that stumped other leading models.”

API access will cost $2 per million input tokens and $12 per million output tokens for prompts with 200,000 tokens or fewer. That’s up from $1.25 and $10 per million Gemini 2.5 Pro input/output tokens.

With this launch, Google is also releasing a client-side bash tool that will allow the model to use shell commands during its agentic workflows so it can navigate the file system, for example. There will also be a hosted bash tool for code generation and prototyping, but for the time being, the hosted tool will only be available in the Gemini API for early access partners.

Deep Think Coming Soon

Google will soon also launch another variant of Gemini 3: Gemini 3 Deep Think. This version will feature an enhanced reasoning model that will likely push the model’s performance even further. Right now, Google is only making this model available to safety testers, but it’ll soon be available to Google AI Ultra subscribers (who pay $250/month) as well.

“Gemini 3 Deep Think outperforms Gemini 3 Pro’s already impressive performance on the benchmarks … like Humanity’s Last Exam and GPQA Diamond. So it really pushes the boundaries of intelligence to help you solve your most complex problems,” Tulsee Doshi, Google’s senior director and head of product for the Gemini models, explained in a press briefing ahead of today’s launch.