It's October 5th 2026, my AI bill is almost free.

If it were 5 months go, I know free AI is just a scam. Because it was literally a scam. You can only get what you pay for! Either you're using a stupid AI, or you're selling your data. Guess what's changed for the past few months? Yep, open-weights models.

Open-weights models have been improving to the point where they're very good at doing a specific field that you choose them for. What do I mean by that? I don't recall but I may have said this in previous posts: You do not to categorize tasks for frontier AI models, You can just use frontier AI models and they can handle disruptive conversation. Which means you can literally ask it 10 questions, be so ambiguous, not matter the size and category of the tasks that it's given, it can handle it very well. But that time is over. It's really over. The frontier AI models today is either too expensive or very dumb. They literally make the good old models dumber so people are forced to use their more expensive ones. You can't really blame them, AI is an industry running on venture capital and on a lot of bets.

So if one's bill of AI ~ $500/month from 6 months ago would quadruple for the same amount of work done today. That was literally my case. Now I don't do that anymore. They're just insane now. But lucky for us, open-weights models are now reaching the point of being very productive and useful if you know how to use it right. And you should learn to make use of them!

You notice that I used past tense in few places? That's because I have a solution ❤️. I value my data and I don't want to hand them out for free. But there are some data that's not that critical and can be public. So what I do is I divide the two and use different routings accordingly. I was surprised on how cheap this approach has been. I subscribed to OpenCode Go for $10 a month and I'm very happy with it. It wasn't very good back then but now with OpenCode v2 and with DeepSeek v4.1 Flash. That's just one example, there are many other open-weights models can perform better than others at different fields, so there wouldn't be any model that outperform everyone else. At least not for so long, others will catch up in no time.

I get very comfortable usage limits for my same workload.

But that's literally handing out data? YES 😂. That's I have self-hosted AI for my own AI model. You have many options to create your own model of your own taste. You can fine-tune, abliterate, shift professions, mix and route models around. You have many options to find out what fits you best. You can have rules for cloud models to handle the main flow and then invoke self-hosted models for private parts and vise versa. You can use Groq for very fast inference, Nvidia Build for free API limits, if you dig deep enough you can find many providers that provide AI access for free (to some extents).

So what do you see the picture I've painted there? You have so many options to-date. And the open-weights model are getting so cheap and so good that you don't need to use the "smartest AI model possible" anymore. Since it's not the smartest that get the job done anymore. You should not think of AI as a human being that works for you. IT IS A TOOL. If you can let that sink in, you will find that using AI is literally very affordable and matches its value + market price for it. Just don't chase for the frontiers anymore.

Last but not least, it's personal, it's just me and my personal use. Corporate use and creative use are different than my use. I highly discourage the use of AI for creative tasks whilst highly encourage the use of AI for "manual" labor of the mind.

You gotta advance, gotta find out what works and willing to change. Oh I forgot about the price for self-hosted one right? I can suggest that you can rent out computes for that. That's really really cheap. So my total bill of AI is about $30.

Tell me what you think. How are you using AI today?