Trying various AIs every month

February 1, 2026

January: Only free options

January begun with the end of last year's Claude subscription - I didn't renew it because I wanted to explore my options better, and it was too easy to burn through the limits when using agents like Claude Code.

I ended up using Google's AI mode a lot, using it as my default search engine. It may appear to be just another tab of Google Search, but it's really fast and powerful. Only downside is the occasional glitches, like how it sometimes refuses or fails to generate a response, or how URLs and code blocks are handled poorly.

If you want to switch to AI Mode, here's how to configure your browser:

• AI Mode (@ai): https://www.google.com/search?udm=50&q=%s
• AI Mode Canvas (@canvas): https://www.google.com/search?udm=50&rc=1&q=%s
• AI Mode Pro (@pro): https://www.google.com/search?aep=1&nem=143&q=%s

I also exhausted almost all 300 premium requests of my Copilot student subscription. This doesn't usually happen, but it did this month, across general chat, refactoring, and agents. It's important to pay attention to the premium request multiplier: I'm trying to switch to Gemini Flash for most of my questions since it's 3x cheaper than most models and 9x cheaper than Opus.

February: ChatGPT Business

When you own a domain, you get lots of free things. One of those things is a month of ChatGPT Business. Unfortunately, even if the models are at the frontier, the UI isn't.

ChatGPT gets a lot of things wrong Claude gets right:

• They haven't had Claude's idea of "just give it a normal computer". Instead, you have to use "Agent mode", and if you want it to have access to GitHub you have to turn on "Developer mode".
• The models aren't trained on aesthetics or planning.
• Thinking is unpredictable, compared to the free plan where when you type /think it thinks. If you select "Thinking", it sometimes chooses to not think or hide what it was thinking about. If you select "Pro" and have a chat, it lets you change the thinking settings, but then doesn't actually use them.

But Codex CLI just works, with generous limits, and demonstrating "less is more" well. Scheduled tasks are neat. ChatGPT Business allows for multiple seats. And ChatGPT excels at all thinking-related tasks - extended thinking, deep research, pro mode, the like.

I ended up barely relying on Copilot, using just 107.61/300 premium requests.

March: ChatGPT Plus

Just as I was about to switch to T3 Chat, ChatGPT gave me an offer: a free month of Plus. I would be stupid to not accept. Another month with GPT.

I'm not actually supporting OpenAI - I'm costing them money - so, anti-military folks, don't cancel me. I've only paid OpenAI when using their API directly (have only paid $66.56 and haven't used since August 2025) and when using their models via GitHub Copilot or Models (I don't do this much).

The most notable loss from Business is Pro. If I want GPT 5.2 Pro, I'm out of luck. The best I get is extended thinking + "think very very hard".

April: GLM

GLM is getting a worse and worse deal. Only 13 coding agents are officially supported. Use one of the 3 other agents and you might face "temporary rate limits". Try to use GLM in another UI and "some subscription benefits may be restricted". And once July 2026 hits, GLM-5.1 will use 2x usage usually and 3x usage during 11PM-3AM PT.

But, when I used it, GLM just worked. Seriously - the only time I had to worry about usage limits was when building a data pipeline from scratch, and the only time I had to worry about censorship was when using Claude Code (should've used OpenCode).

GLM can't do "grind to exact goal" like GPT can. It won't catch bugs just by thinking. But it's obviously Sonnet-level and can do some amazing things if you give it a good environment and plenty of guidance.

May: Only "free" options

First: mobile. ~All of my AI usage on my phone switched to this AI Mode quick launcher. AI Mode is very speedy, and this launcher just made it much quicker to access. I regret not doing this last month.

Then the tumultuous part - what happened in agents this month.

• I still had access to Copilot, but GitHub removed nearly all Claudes and all new GPTs, and was on the verge of switching to usage-based billing instead of request-based.

• I was grandfathered in to CrofAI. CrofAI is really good - the models can do things like autonomously debug a problem and use git + gh to submit a fix.

  And for a while, nearly anyone could get the same experience by buying a plan, getting a certain number of daily requests at a flat monthly rate. But this was unsustainable, so as of May 31 CrofAI dropped plans in favor of pure pay-as-you-go. CrofAI's pay-as-you-go prices are really good and how I found them in the first place, and if I moved my workload to some of the models with low prices for cached input (DeepSeek, MiMo) I'd be around the $10-20 range.

• And then a friend found a source of inference on frontier models that I have to hold myself back from naming because it's so [filler] funny. Sorry. Gatekeeping, tragedy of the commons, some of your friends are liars, some more of your friends are gossipers, all that.