My evaluation of AI agent use vs. enjoying life as a human

First, to be clear: except for writing technical books and providing free mentoring and volunteer services I am retired so my experience may not be relevant to yours.

I subscribed to Gemini ultra which is a $250 a month package and did a wide variety of software development task using both Gemini ultra and model is like DeepSeek v4 with OpenCode. For the first two weeks most test projects took half as long when I carefully wrote an initial one page prompt with my goals, what I wanted researched and started Antigravity+Claude Opus and followed up with periodic interactions. I used the long breaks to get up from my desk, do short Qi Gong exercises, etc.

When I used this same working style with DeepSeek v4 (and other frontier open models) it would take twice as long to get a project completed to my satisfaction.

Then I decided that I was using the same work flow for fundamentally different tools. I kept my process the same for Antigravity+Claude Opus but took a different approach using open models and open source tooling:

Now, I don’t always use DeepSeek v4 Pro and instead for small easy tasks I use DeepSeek v4 Flash that is so fast that there are no real pauses in an interactive workflow. Even when I switch to DeepSeek v4 Pro that is just slight slower (I buy API credits from both Fireworks.ai and DeepSeek Platform), I use a more interactive collaborative work flow in which I do more of the work and primary decision making.

Comparing how I feel after work sessions

I find that after using higher end agentic tools like Antigravity+Claude Opus that I feel ‘fuzzy headed’ or tired in weird ways. Frankly I don’t understand this since I am not working as hard, but that is how I feel.

When I have more skin in the game and am actively editing code and prompts in emacs with OpenCode + DeepSeek v4 (and sometimes I like Claude Code configured for DeepSeek v4) I feel vitalized and in a way more refreshed when work in this mode.

Cancelled Gemini and for a few months I am going to go with a more interactive work flow with open models

I don’t believe in getting locked into tools but there is efficiency with sticking with a set of tooling for a while. I cancelled my Gemini subscription a few days ago.

Whoa there, what about the environment?

I have no doubt that using slightly weaker but much more efficient models uses less energy, compute, water, and is easier on our fragile environment. I see people trying to argue the opposite, but they are either being foolish or have some agenda they are pushing.

I see real hope in more efficient models: I have spent about five hours evaluating the combination of Gemini 3.5 Flash with Antigravity CLI: great combination but since I am not paying for tokens (when using a subscription) it seems like comparing Apples and Oranges evaluating against DeepSeek v4 where I clearly see the dollar cost (or actually ‘pennies cost’) for small individual tasks. I will try to evaluate Gemini 3.5 Flash with Antigravity CLI paying for tokens vs. DeepSeek v4 + Claude Code (always paying for tokens with this option) sometime in the next few months and edit this article with that result.