I use LLMs for code at work, but I've been a bit hesitant to dive in for side projects because I'm worried about the cost. Is it necessary to pay $200/mo to actually ship things or will $20/mo do it? Obviously I could just try it myself and see how far I get bit I'm curious to hear from someone a bit further down the path.
I have a feeling you are using SOTA models at work and aren't used to just how cheap the non-Anthropic/Google/OAI options are these days. GLM's coding subscription is like $6/month if you buy a full year.
Depending on how much you use, you can pay API prices and get pretty far for 20 bucks a month or less. If you exhaust that, surprisingly, I recommend getting Gemini with the Google AI pro subscription. You can use a lot of the Gemini CLi for that
The $20/mo subscription (Claude Code) that I've been using for my side projects has been more than enough for me 90% of the time. I mostly use the cheaper models lately (Haiku) and accept that it'll need a bit more intervention, but it's for personal stuff and fun so that's ok. If you use VSCode, Antigravity or another IDE that's trying to market their LLM integration, then you'll also get a tiny allowance of additional tokens through them. I'll use it for a few hours at a time, a couple days a week, often while watching TV or whatever. I do side projects more on long rainy weekends, and maybe not even every week during the summer. I'll hit the limit if I'm stuck inside on a boring Sunday and have an idea in my head I really wanted to try out and not stop until I'm done, but usually I never hit the limit. I don't think I've hit the limit since I switched my default to Haiku FWIW. The stat's say I've generated 182,661 output tokens in the last month (across 16 days), and total usage if via API would cost $39.67.
They all let you do that now, including Claude Code itself. You can choose between pay per token and subscription. Which means that a sensible way to go about those things is to start with a $20 subscription to get access to the best models, and then look at your extra per-token expenses and whether they justify that $200 monthly.
Wait till the VC tap gets shut off. You: Hey ChatGPT, help me center a div. ChatGPT: Certainly, I'd be glad to help! But first you must drink a verification can to proceed. Or: ChatGPT: I'm sorry, you appear to be asking a development-related question, which your current plan does not support. Would you like me to enable "Dev Mode" for an additional $200/month? Drink a verification can to accept charges.
I can absolutely see that happening. It's already kind of happened to me a couple of times when I found myself offline and was still trying to work on my local app. Like any addiction, I expect it to cost me some money in the future
This is currently negative expected value over the lifetime of any hardware you can buy today at a reasonable price, which is basically a monster Mac - or several - until Apple folds and rises the price due to RAM shortages.
The next best thing is to use the leading open source/open weights models for free or for pennies on OpenRouter [1] or Huggingface [2]. An article about the best open weight models, including Qwen and Kimi K2 [3]. [1]: https://openrouter.ai/models [2]: https://huggingface.co [3]: https://simonwillison.net/2025/Jul/30/
This requires hardware in the tens of thousands of dollars (if we want the tokens spit out at a reasonable pace). Maybe in 3-5 years this will work on consumer hardware at speed, but not in the immediate term.
$2000 will get you 30~50 tokens/s on perfectly usable quantization levels (Q4-Q5), taken from any one among the top 5 best open weights MoE models. That's not half bad and will only get better!
If you are running lightweight models like deepseek 32B. But anything more and it’ll drop. Also, costs have risen a lot in the last month for RAM and AI adjacent hardware. It’s definitely not 2k for the rig needed for 50 tokens a second
It's amazing to be able to try a bunch of ideas with very minimal cost. That being said, AI code assistants don't have eyeballs and they often make things that don't look very good. Craft, polish and judgement still matter.
And even more fun with tools/services like exe.dev! Also apparently the combined of Google Antigravity/$20 Google AI plan/Opus 4.5 is blowing up the AI community lately in Reddit. Apparently the limits right now of Opus thru Antigravity are insanely generous/incredible value. Obviously this could change at any time but perhaps Google has the funds/resources to continue to provide value like this in an attempt to capture the dev userbase / win the AI war.
Related question which might fit here so I'm going to try: What is the absolute cheapest way to get started on AI coding a simple website? I have a couple ideas I want to test out and get out of my head and onto the web but have resisted for years because my webdev knowledge is stuck in 2004 and I've had no desire to change that. These are not complicated things (all static, I think) but... I hate webdev. I am not really willing to pay to do any initial explorations, but if I like where things are going then, sure, I'll pay up. I have a decently powerful machine that can run things locally, but it is Windows (because I'm an EE, sadly), which does matter.
Google Gemini has a generous free tier. You could start by experimenting in AI Studio - https://aistudio.google.com/ - then have a go at coding agents using their Gemini CLI or Antigravity tools. For what you're describing the free tiers of the Claude and ChatGPT web apps would probably work fine too.
And now you can generate javascript infested slop frameworks for $5 per million tokens. Such an improvement. And it's so easy to just ask Claude to make one for you, why even bother standardizing anything when you can just use bespoke slop for anything anymore. Libraries and frameworks? Not needed. Just shove everything into CC/Codex and let it figure it out.