Gemini CLI

The killer feature of Claude Code is that you can just pay for Max and not worry about API billing. It lets me use it pretty much all the time without stressing over every penny or checking the billing page. Until they do that - I'm sticking with Claude.

How does that compare to using aider with Claude models?

That's a golden cage and you limit yourself to Anthropic only.

I'm happy I can switch models as I like with Aider. The top models from different companies see different things in my experiences and have their own strengths and weaknesses. I also do not see Anthropic's models on the top of my (subjective) list.

No per-token billing here either: "...we offer the industry’s largest allowance: 60 model requests per minute and 1,000 requests per day at no charge."

https://blog.google/technology/developers/introducing-gemini...

Same. Generally i really prefer Claude Code's UX (CLI based, permissions, etc) - it's all generally close to right for me, but not perfect.

However i didn't use Claude Code before the Max plan because i just fret about some untrusted AI going ham on some stupid logic and burning credits.

If it's dumb on Max i don't mind, just some time wasted. If it's dumb on credits, i just paid for throw away work. Mentally it's just too much overhead for me as i end up worrying about Claude's journey, not just the destination. And the journey is often really bad, even for Claude.

This insistence by SAAS vendors upon not protecting you from financial ruin must surely be some sort of deadweight loss.

Sure you might make a few quick wins from careless users but overall it creates an environment of distrust where users are watching their pennies and lots are even just standing off.

I can accept that with all the different moving parts this may be a trickier problem than a pre paid pump, or even a Telco, and while to a product manager this might look like a lot of work/money for something that “prevents” users overspending.

But we all know that’s shortsighted and stupid and its the kind of thinking that broadly signals more competition is required.

I did a little digging into this just yesterday. The impression I got was that Claude Code was pretty great, but also used a _lot_ more tokens than similar work using aider. Conversations I saw stated 5-10x more.

So yes with Claude Code you can grab the Max plan and not worry too much about usage. With Aider you'll be paying per API call, but it will cost quite a bit less than the similar work if using Claude Code in API-mode.

I concluded that – for me – Claude Code _may_ give me better results, but Aider will likely be cheaper than Claude Code in either API-mode or subscription-mode. Also I like that I really can fill up the aider context window if I want to, and I'm in control of that.

If you use your personal gmail account without billing enabled, you get generous requests and never have to worry about a surprise bill.

Using Claude models in aider burns tokens you need to top up. With Claude Max subscription you can pay a 100 or 200 USD per month plan and use their internal tool claude code without the need to buy additional pay as you go tokens. You get a "flatrate", the higher plan gives you more usage with less rate limiting.

> I concluded that – for me – Claude Code _may_ give me better results, but Aider will likely be cheaper than Claude Code in either API-mode or subscription-mode.

I'd be pretty surprised if that was the case - something like ~8 hours of Aider use against Claude can spend $20, which is how much Claude Pro costs.

If I have a CC linked to my personal Google for my Google One storage and YouTube Premium, that doesn't make me "billing enabled" for Gemini CLI does it?

Don’t know about Claude, but usually Google’s free offers have no privacy protections whatsoever — all data is kept and used for training purposes, including manual human review.

Indeed, I think I came to the incorrect conclusion! Just signed up for a subscription after getting through quite a lot of API funds!

Aider and Claude Code/Gemini CLI agentic stuff operate differently.

You can think of Aider as being a semi-auto LLM process. First you ask it to do something. It goes through a generate -> reflect -> refine loop until it feels like it has achieved the goal you give it. Aider has a reflection limit so it'll only do this loop a limited number of times and then it will add/remove the code that it deems fit. Then it'll give you instructions to run. You can run those instructions (e.g. to actually run a script) and then append the results from the run into the context to get it to fix any issues, but this is optional. What you send in the context and what you ask the models to do are in your hands. This makes iteration slower and the LLM does less but it also can potentially keep costs lower depending on what you delegate to the LLM and how often you iterate with it.

Claude Code, Codex, and I suspect Gemini CLI on the other hand will autonomously run your code then use the output to continue refining its approach autonomously until the goal is reached. This can consume many more tokens, potentially, than hand guiding Aider, because its potential for iteration is so much longer. But Claude Code and the like also need a lot less direction to make progress. You can, for example, ask it to do a big refactor and then just leave to lunch and come back to see if the refactor is done. Aider will require babying the whole way.