Google Gemini 3.6-3.8 Impressions

Randoms thoughts and impressions from using the Gemini 3.6-3.8 models. No benchmarks were used. This has just been from my personal experience.

Gemini Flash 3.6-3.8

TL:DR

Pros:

• As a generalist chat bot and researcher, It is better than anything else I have tried
• It is fast. Really really fast
• Superior prose editor
• Superior front end designer
• The $20/month Google AI Pro is the best value, I've found.

Cons:

• It is a weaker backend software developer than its peers from other moel labs.
• The antigravity-cli is awful and a downgrade from the gemini-cli
• Using Gemini outside of antigravity-cli while on the Google AI Pro plan technically is a breach of the Terms & Conditions which is annoying.

DevOps Work

Yeah I'm grouping these 3 together since they basically all came out at the same time. 3.8 ended up being the main driver for my first run through of setting up all of the different open source applications on this kubernetes cluster. It was quite good about the initial setup and I am honestly quite pleased with the results I got. Yes, it hardcoded some secrets and messed up actually getting the apps into a gitops delivery pattern even after ArgoCD was setup but overall I was pretty happy with the results. I was using the antigravity-cli (agy) as the harness since the google AI Pro plan frowns if you use it in non-antigravity harnesses. I don't particularly like the agy harness and think its a significant downgrade from the previous gemini-cli harness but Google randomly abandoning software is nothing new.

Chat Bot and Researcher

As a general chat and research bot, this is my main driver. You sometimes need to explicitly prompt it to actually do web research but I've noticed over time that their models have become alot better about defaulting to web_search (or whatever the tool is name is for them) rather than defaulting to their own training corpus. "Who is the starting right guard for the Philadelphia Eagles?" was returing the correct answer for 2024 which is definitely wrong when you ask it that question in April 2026. But it has gotten better there. A real blind spot for it though is actually asking about new LLM models or "the best models that fit" a specific GPU... a lot of Qwen2.5 and Gemma3 recommendations which are ... several generations behind at this point. But these are issues that all the chat bots from all the vendors have and the Gemini models seem better than the others.

Also, Gemini Notebook (formerly LLM Notebook) is just an awesome tool for web research and summarization and I wonder why Microsoft Copilot doesn't have feature parity with this.

Prose Writer

Unlike models from other competing frontier labs (from both US and China), the Gemini model don't seem to suffer from the same awful AI "voice" as the others.

• They don't overuse em dashes (-).
• None of the "Hey I found 3 issues and 2 that really matter" nonsense
• They are less aggressive or authoritative sounding in their recommendations (although not completely without issues)

As such, I have generally gravitated to the Gemini models when I need something to help edit my own prose; a meandering but blunt style that doens't quite lend itself well to corporate speak.

Frontend Design

The Gemini models are just better than the other models full stop. Aesthetically, the frontend designs that come from Gemini are more fleshed out and professional looking than the ones I have produced from using Anthropic. The Anthropic models just default to the same tired look over and over again regardless of how much I have tried to correct for it.