Skip to main content
  • Pricing
News

Gemini 3 Pro One Year On: What Google Claimed at Launch and What Happened Next

By Oakgen TeamPublished 3 min read

Share
Gemini 3 Pro One Year On: What Google Claimed at Launch and What Happened Next

Gemini 3 Pro is the Google model announced on November 18, 2025 and released in preview that day. As of November 19, 2026, the original gemini-3-pro-preview endpoint has been shut down, and Google's documentation points developers to newer models. This guide covers what Google claimed at launch, the published limits, and how to read those claims a year later.

What Google announced

According to Google's launch post, Gemini 3 Pro was released in preview and made available across Google products, including the Gemini app, AI Studio and Vertex AI. Google called it its most intelligent model and said it significantly outperformed Gemini 2.5 Pro on every major benchmark. A Deep Think reasoning mode was announced alongside it, with access first going to safety testers before Google AI Ultra subscribers. Google also introduced Antigravity, an agentic development platform, on the same day.

The benchmark numbers Google published

These figures come from Google's own post, so treat them as vendor-reported results rather than independent measurements.

BenchmarkGemini 3 Pro (Google's figure)What it measures
LMArena1501 EloHuman preference votes
WebDev Arena1487 EloWeb development tasks
Humanity's Last Exam (no tools)37.5%Academic reasoning
GPQA Diamond91.9%Graduate-level science questions
MathArena Apex23.4%Hard mathematics
MMMU-Pro81%Multimodal reasoning
Video-MMMU87.6%Video understanding
SimpleQA Verified72.1%Factual accuracy
Terminal-Bench 2.054.2%Operating a computer through a terminal
SWE-bench Verified76.2%Coding agents

For Deep Think, Google reported 41.0% on Humanity's Last Exam without tools, 93.8% on GPQA Diamond, and 45.1% on ARC-AGI-2 with code execution.

Published limits

Google's model card describes Gemini 3 Pro as a sparse mixture-of-experts transformer with native multimodal support. It takes text, images, audio and video as input with a context window of up to 1 million tokens, and outputs text up to 64K tokens. The model card gives a knowledge cutoff of January 2025 and lists hallucinations among the known limitations, along with occasional slowness or timeouts. The developer guide recommends Search Grounding for information newer than the cutoff.

Where it stands now

The Gemini API deprecations page lists gemini-3-pro-preview with a release date of November 18, 2025, a shutdown date of March 9, 2026, and gemini-3.1-pro-preview as the recommended replacement. Google's model page for the old endpoint now carries a warning that it was shut down on March 9, 2026, and tells developers to migrate to Gemini 4 Argon.

The models that followed:

  • Gemini 3.1 Pro. The deprecations page lists gemini-3.1-pro-preview with a release date of February 19, 2026 and no announced shutdown date. Google's developer guide lists it with a 1M input and 64k output window and pricing of $2 input and $12 output per million tokens for prompts under 200k tokens, and $4 and $18 above that. The guide says there is no free tier for it in the Gemini API.
  • Gemini 4 Argon. Google's model page lists gemini-4-argon as stable, with a unified context window of 1,048,576 tokens and a default maximum output of 65,536 tokens, extendable to 1,048,576 with long decode.

On Google DeepMind's Gemini 3.1 Pro page, Google's own comparison table puts Gemini 3.1 Pro ahead of Gemini 3 Pro on several benchmarks, for example Humanity's Last Exam at 44.4% against 37.5%, ARC-AGI-2 at 77.1% against 31.1%, and SWE-Bench Verified at 80.6% against 76.2%. That table is again Google's, and the page labels 3.1 Pro as a preview.

How to read launch-day claims

Three habits help when you read any model launch:

  1. Check who measured it. A vendor blog post reports vendor-run results. Independent leaderboards and your own tests answer different questions.
  2. Check the date. A leaderboard rank is true on one day. The launch post's "tops the leaderboard" statement describes November 2025, not today.
  3. Check the endpoint. A model name in an article may map to an endpoint that has since been retired. The deprecations page is the place to confirm.

Testing a model on your own work

Benchmarks rarely match your use case. A practical approach is to write five to ten prompts from your real work, such as a document summary, a spreadsheet formula, a bug report and an email draft, and run the same set on two models. Score each answer on accuracy, completeness and tone, and keep the prompts so you can rerun them when a new version appears.

Take this further with Oakgen Agent

Write a test-prompt set and scoring rubric, explain how to run them, and create an xlsx or csv scoring sheet; can search the web for current model details.

Open this task in Oakgen Agent

The prompt is filled in for you. Nothing is sent until you press send.

Open in Oakgen Agent

Frequently asked questions

When was Gemini 3 Pro released?

Google announced Gemini 3 on November 18, 2025 and released Gemini 3 Pro in preview across its products, including the Gemini app, Google AI Studio and Vertex AI. The Gemini API deprecations page lists gemini-3-pro-preview with a release date of November 18, 2025.

Is Gemini 3 Pro still available in the Gemini API?

As of 2026-11-19, no. Google's model page for gemini-3-pro-preview says it was deprecated and shut down on March 9, 2026, and the deprecations page lists gemini-3.1-pro-preview as the recommended replacement. The same model page now points developers to Gemini 4 Argon.

What context window and output limit does Gemini 3 Pro have?

Google's model card for Gemini 3 Pro says it accepts text, images, audio and video with a token context window of up to 1 million tokens and produces text output of up to 64K tokens. The model card gives a knowledge cutoff of January 2025.

Did Gemini 3 Pro really rank first on leaderboards?

At launch Google reported that Gemini 3 Pro topped the LMArena leaderboard with a score of 1501 Elo and the WebDev Arena leaderboard with 1487 Elo. Those are Google's statements from the launch date. Leaderboards change as new models are added, so they do not tell you the current ranking.

Sources

  1. Google, A new era of intelligence with Gemini 3, Nov 18, 2025 (read 2026-11-19)
  2. Google DeepMind, Gemini 3 Pro model card, last updated May 2026 (read 2026-11-19)
  3. Google AI for Developers, Gemini deprecations (read 2026-11-19)
  4. Google AI for Developers, Gemini 3 Pro preview model page (read 2026-11-19)
  5. Google AI for Developers, Gemini 3 developer guide (read 2026-11-19)
  6. Google DeepMind, Gemini 3.1 Pro model page (read 2026-11-19)
  7. Google AI for Developers, Gemini 4 Argon model page (read 2026-11-19)
Share