Google Gemini 3 release date, features and what's new — infographic

Google Gemini 3: Release Date, Features, and What’s New

ByAIERA Blog Editorial Team

Key Takeaways

  • Google Gemini 3 was announced on November 18, 2025, with Google calling it its “most intelligent model” and launching Gemini 3 Pro in preview.
  • Headline capabilities include generative UI (the model builds interactive experiences, not just text answers), 76.2% on SWE-bench Verified, 1501 on LMArena, 91.9% on GPQA Diamond, a Deep Think long-reasoning mode for AI Ultra subscribers, and the Antigravity agentic coding platform.
  • On the imaging side: Nano Banana Pro launched November 20, 2025, built on Gemini 3 Pro, and the gemini-3-pro-image API model followed on May 28, 2026.
  • New in October 2026: the “Gemini agent,” unveiled at the Gemini at Work 2026 event on October 8 — a universal agent running on Gemini 3 Pro that writes and runs code, coordinates sub-agents, and integrates with Workspace apps plus Slack and Microsoft 365. It’s in private preview with pricing undisclosed.
  • Preview API pricing started at $2 input / $12 output per million tokens (up to 200K context), rising to $4/$18 above that; consumers can access it through the $20/month Pro tier.

Introduction

Is Google Gemini 3 actually out, and should you care? The short answer: yes, it’s been out since November 2025 — but Google’s naming, the parade of variants, and a wave of newer announcements have made the picture confusing. This guide pulls together what’s confirmed: the original launch, the benchmark numbers Google reported, the image-generation models built on it, and the brand-new Gemini agent that runs on Gemini 3 Pro. Where details are still unconfirmed, we’ll say so.

Google Gemini 3: The Announcement in Brief

Google Gemini 3 was announced on November 18, 2025, with Gemini 3 Pro available in preview from day one.

Since then, the “Gemini 3” name has covered a growing family: Gemini 3 Flash and later 3.x variants arrived through API releases, and preview aliases have changed over time. That matters in practice — a family name is not the same as an API endpoint, so developers should always check the exact model ID and its lifecycle stage before building on it. A community-maintained release tracker has been useful for keeping the variants straight.

Key Capabilities and Benchmark Numbers

These are the figures Google reported at launch (all from the company’s own materials, not independent verification):

  • Reasoning: 91.9% on GPQA Diamond, a benchmark of PhD-level science questions, and 37.5% on Humanity’s Last Exam.
  • General capability: 1501 on LMArena, surpassing its predecessor’s 1451 at the time.
  • Coding: 76.2% on SWE-bench Verified — real-world software engineering tasks — which Google framed as making it the company’s best coding model to date.
  • Deep Think mode: an extended-reasoning mode for complex problems, available to Google AI Ultra subscribers.
  • Google Antigravity: a new agentic development platform for autonomous, task-oriented software development, launched alongside the model.

The honest caveat: launch-day benchmarks are chosen by the vendor. They tell you a model is competitive, not that it wins every workload. If you’re picking a model for production work, run your own evaluations on your own tasks.

Generative UI: The Standout Feature

The most distinctive idea in Gemini 3 is what Google calls generative UI (or generative interfaces). Instead of always answering in paragraphs, the model decides the best format for its response — generating an interactive web-like interface, a custom calculator, or a physics simulation on the fly, tailored to your question.

Ask for travel recommendations and you might get a visual, interactive layout with follow-up questions rather than a wall of text. It’s a genuine shift in how humans interact with AI: from a chatbot that answers to a system that builds experiences. Whether it proves more useful than annoying in daily use depends on execution, and it’s an area where hands-on testing beats marketing claims.

Image Generation: Nano Banana Pro and gemini-3-pro-image

Two image models sit on the Gemini 3 Pro foundation:

  • Nano Banana Pro, announced November 20, 2025 — Google’s advanced image-generation model built on Gemini 3 Pro.
  • gemini-3-pro-image, released via API on May 28, 2026.

One caution from recent history: Google moves fast on model lifecycles. Mixed News reported that a related Gemini image model was retired just 154 days after release, on 23 days’ notice. Google’s deprecation policy promises advance notice but doesn’t fix a minimum period — so if you build products on these endpoints, watch the deprecation page and have a migration plan.

The New Gemini Agent (October 2026)

The freshest Gemini 3 news came on October 8, 2026, at the Gemini at Work 2026 event, where Google unveiled the “Gemini agent” — a universal workplace agent running on Gemini 3 Pro, covered by Reuters and CNBC.

What it does, per launch reporting: answer questions, write and run code, create documents and media, and coordinate sub-agents that can work autonomously for hours or even days on an objective. It plugs natively into Gmail, Docs, Drive, Sheets, and Calendar, with third-party integrations for Slack and Microsoft 365. It’s currently in private preview for select customers through the Gemini Enterprise app and Workspace, and pricing was not disclosed at launch.

An important limitation noted in the launch coverage: no benchmark numbers were published tying the agent’s coding ability to competitors like GPT-6 or Claude, so claims about coding superiority are marketing until measured. Treat the agent as promising but unproven until independent testing exists.

Pricing and Availability

Where things stand on access and cost:

  • API (preview pricing): $2 per million input tokens / $12 per million output tokens for up to 200K tokens of context; $4/$18 above 200K.
  • Consumers: available through the $20/month Pro subscription, with a free tier in Google AI Studio.
  • Enterprise: Gemini 3 powers features in Google’s enterprise products, AI Mode, and AI Overviews.

How It Compares — Honestly

Gemini 3’s reported numbers put it firmly in the frontier tier alongside OpenAI’s GPT-5 series and Anthropic’s Claude models. But head-to-head superiority claims should be treated carefully: vendors benchmark on different tasks, with different prompting, at different times. For a deeper look at one of its main rivals, see our guide to DeepSeek AI, and for the broader Gemini lineup, our Gemini explainer.

Trying It Yourself

The lowest-friction way to test Google Gemini 3 is the free tier in Google AI Studio, which lets you experiment with prompts, the generative UI features, and the API without paying. Developers evaluating it seriously should pin an exact model ID rather than a “latest” alias, record the date tested, and run the same task set against one or two competing models — vendor benchmarks are a starting point, not a verdict.

What We Still Don’t Know

  • Independent verification of the newest benchmark claims, especially for the October 2026 Gemini agent.
  • Agent pricing and general availability — both undisclosed at launch.
  • Long-term model lifecycle — with variants and aliases shifting, the exact endpoint you build on today may not be the one serving you next year.

Bottom line: Google Gemini 3 is a real, shipping model family with genuinely interesting ideas (generative UI most of all), strong reported benchmarks, and a growing agent ecosystem around it. Just verify the specific endpoint, price, and availability for your use case before committing.

Similar Posts