The Memo - Special edition - Gemini 3 - Nov/2025

The best frontier model in the world (for 2025)...

Nov 18, 2025

To:      US Govt, major govts, Microsoft, Apple, NVIDIA, Alphabet, Amazon, Meta, Tesla, Citi, Tencent, IBM, & 10,000+ more recipients…
From:    Dr Alan D. Thompson <LifeArchitect.ai>
Sent:    18/Nov/2025
Subject: The Memo - AI that matters, as it happens, in plain English
AGI:     95%
ASI:     0/50 (no expected movement until post-AGI)

Google announces Gemini 3

Once again, we have this out to The Memo readers within just a few hours of model release.

This is the one we’ve been waiting for. This is what GPT-5 should have been. Google said:

Gemini 1’s breakthroughs in native multimodality and long context window expanded the kinds of information that could be processed — and how much of it. Gemini 2 laid the foundation for agentic capabilities and pushed the frontiers on reasoning and thinking, helping with more complex tasks and ideas, leading to Gemini 2.5 Pro topping LMArena for over six months.
And now we’re introducing Gemini 3, our most intelligent model, that combines all of Gemini’s capabilities together so you can bring any idea to life.
It’s state-of-the-art in reasoning, built to grasp depth and nuance — whether it’s perceiving the subtle clues in a creative idea, or peeling apart the overlapping layers of a difficult problem. Gemini 3 is also much better at figuring out the context and intent behind your request, so you get what you need with less prompting.

Watch the video (link):

Size estimates

General model size is no longer an indicator of performance, but I still find it interesting. With all model details kept confidential, plus added complexity in reasoning/thinking mode, it is more challenging than ever to estimate token and parameter counts.

Now in 2025, based on my ongoing analysis, known Gemini 3 model pricing, similar known frontier MoE model sizes and pricing*, estimates of training supply (TPUs), inference supply (TPUs), and demand (users), here are my initial estimates for the Gemini 3 model.

* Grok-3, Grok-4, and Grok-5 were the first major frontier models to have their sizes publicly detailed, with the CEO of xAI recently disclosing (15/Nov/2025): ‘[Grok-5] is a 6 trillion parameter model, whereas Grok-3 and -4 are based on a 3 trillion parameter model.’

For dataset size, Gemini 3 Pro was likely trained on 100T tokens seen. Data cut-off is Jan/2025.

For parameters, I estimate Gemini 3 Pro to have a centrepoint of around 3T parameters MoE.

Benchmark scores

Gemini 3 scores GPQA=93.8 (with Deep Think), HLE=45.8 (with tools search + code exec). This is still just slightly below my criteria for an ASI system, so we’ll have to wait for the next frontier model to hit that indicator. (‘It is likely that any model with a primary score at >50% on HLE, and a secondary score at >90% on GPQA is an ASI system.’ https://lifearchitect.ai/asi/)

https://lifearchitect.ai/models-table#rankings Click to enlarge.

https://blog.google/products/gemini/gemini-3/#gemini-3 Click to enlarge.

I’ve run my ALPrompt benchmark scores across Gemini 3, and the results are excellent, with the current high score across all prompts (including the new shortest word for 2025H1):

Gemini 3 ALPrompt 2025H2 score: 2/5 (GPT-5=2/5)
Gemini 3 ALPrompt 2025H1 score: 4/5 (GPT-5=3/5)
Gemini 3 ALPrompt 2024H2 score: 5/5 (GPT-5=5/5)
Gemini 3 ALPrompt 2024H1 score: 5/5 (GPT-5=5/5)

Gemini 3 also set a new record on the Extended NYT Connections=96.8. (Next best is Grok-4=92.1) Read more: https://github.com/lechmazur/nyt-connections/

Also, Dr Jeff Dean (18/Nov/2025) said ‘Each of the games shown in the video below was created by Gemini 3 from scratch w/a few natural language prompts/images.’ Check out the ‘playables’ games on YouTube: https://www.youtube.com/@googledeepmind/playables