Why Are AI Coding Tools Getting More Expensive, and What Can You Do About It?
Just when everyone is getting the hang of vibe coding, it feels like the AI companies are trying to pull the rug from under us. In mid-April, GitHub announced on its website that it would suspend new sign-ups for its Copilot subscription plans. Further, they are setting tighter usage limits.
Site owner may receive compensation for purchases made through the links on this website.
In case you’re not aware, vibe coding is conversational software development in which humans guide intent, and AI accelerates implementation through rapid, iterative collaboration.
Will other AI coding assistants follow suit?
AI is not getting any cheaper. In fact, with many LLM companies bleeding cash, these rate hikes were inevitable. While these developments are not the most welcome news, it’s unlikely they spell the end for vibe coding. We just need to adjust, and that is the reason for this article.
[Source: https://taptwicedigital.com/stats/openai]Did you know that several local models can be used for Vibe coding? However, if you don't have the right kind of machine, local models won't perform well (slow). These same local models can be set up to run on RunPod, though. Find out how!
TL;DR Version
The article argues that AI coding tools are becoming more expensive because “agentic” AI workflows consume far more compute and tokens than earlier autocomplete-style coding assistants. Companies like GitHub are tightening limits because users are exhausting infrastructure faster than expected, especially with full-project conversational coding workflows.
The core takeaway is that vibe coding is not dying, but developers need to adapt:
- Use multiple AI models/providers to avoid exhausting weekly token limits.
- Maintain strong project documentation (markdown files, architecture notes, TODOs) so you can switch models seamlessly.
- Use local models through Ollama when possible, especially for lower-cost experimentation.
- Consider cloud GPU rentals, such as RunPod, if your local hardware is weak.
- Time your usage around reset periods and conserve premium credits for complex tasks.
- Open-source and smaller coding models are improving rapidly and may become viable lower-cost alternatives to frontier models.
The article also pushes back on the idea that higher AI costs will necessarily raise developer salaries. Instead, it suggests that the more likely outcome is a hybrid ecosystem in which developers combine premium frontier models with cheaper open-source tools.
FULL VERSION
My Personal Experience
I’ve been really leaning into vibe coding lately. Yesterday, I noticed that the free credits for GitHub’s Copilot were 50% consumed. I hardly used them in the past several weeks, and I started wondering how I can be at the 50% mark. That’s what prompted me to investigate (and also led me to write this).
So, What’s Going On?
Before we discuss what to do, we need to understand why AI companies are potentially raising their rates or lowering their usage limits. Agentic AI is believed to be the main culprit. Consider that when GitHub Copilot emerged, agentic workflows were not mainstream, if they even existed.
GitHub was one of the earliest adopters of AI-assisted coding, but the coding craze really took off when Cursor was introduced alongside Anthropic’s Sonnet. This combination enabled developers to work conversationally with AI across entire projects, rather than just getting line-by-line suggestions.
Agentic workflows can quickly deplete credits, and many providers underestimated how quickly users would adopt them. As stated by GitHub on the page describing the changes:
“As Copilot’s agentic capabilities have expanded rapidly, agents are doing more work, and more customers are hitting usage limits designed to maintain service reliability. Without further action, service quality degrades for everyone.”
[Source: https://github.blog/news-insights/company-news/changes-to-github-copilot-individual-plans/]How Users Feel About This?
Users are divided on how they feel about these changes, but many are left wondering whether they can afford to continue working with GitHub. On Dev.to’s blog, one person sums it up as follows:
GitHub Copilot was successful because it made AI assistance feel integrated, simple, and worth paying for.
“These changes move it in the opposite direction. The product is becoming harder to trust, harder to budget, and harder to recommend.”
[Source: https://dev.to/mvm/github-copilots-pricing-changes-arent-just-expensive-theyre-a-trust-problem-h6g]One YouTuber, in his video (below), argues that these changes, should they spread to other companies, would be good for developers. He says that forcing people to pay more will equalize developers’ salaries with those in the U.S., because it will squeeze out amateurs and junior developers. He also made other points, which you can view for yourself if you’re so inclined to watch the video.
[https://www.youtube.com/watch?v=js3MvXF2Aew]As for price rising causing higher salaries, anything is possible, I suppose. But theory has many unknowns. Mainstream AI use is still in its infancy, especially vibe coding. Even experts don’t know the true impact of where all of this is heading. If you don’t believe that, notable experts often have disagreements about AI’s future. I am not claiming I know either, by the way.
What I do know, however, is what’s already here. I think a more likely scenario is that capable open-source models are already here, and low-cost specialized (can you say coding?) Small Language Models (SLMs) exist or are currently being developed. I did a preliminary experiment on a 30b qwen model, and it was able to read and update my code.
It certainly wasn’t as good as Codex or Claude. But it will be enough for someone to learn the fundamentals of vibe coding.
What Approach to Take Now
- Use multiple models. If you rely on only a few models and you are logging in many hours of vibe coding time, you almost assuredly will exhaust your weekly token allowance way before the week is up. And if the powers that be are becoming more stingy with their allocations, it will only exacerbate the problem. Using multiple models (and some that have unlimited tokens – more on that below) will drastically reduce the occurrence of running out.
- Initially, have the model create markdown files for documentation. Without getting into too many details here (plenty of resources on YouTube and the web), you want the LLM to scan through all the code and create documents about the project, what the objectives are, the current state of the coding, the architecture, a to-do list, any coding standards, plus any other items that you’d believe should be tracked. You can use the AI (whichever model you choose) to help.
- Have the AI update the documents based on all the changes from your session (at the end). When you switch models, these documents will serve as the project’s communication channel. Some models will update them automatically, but it can’t hurt to have them run through the updates at the end of your session. There may be better ways to do this, but it’s one that has definitely worked for me with little interruption.
- Use local models. Ollama (from Ollama.com) is an open-source solution that lets you run models locally on your machine. Anyone can download the software, but unless you have a decent GPU card, you may find working with Ollama slow. You can run the smaller models on most PCs even without a GPU. But as you try to scale up (and you’ll likely need to for this solution to work on vibecoding), running larger models is excruciatingly slow and may even time out.
- Use cloud models. The cloud models listed here are a small sample of the ones available. Be sure to check frequently for others. Here are common cloud models:
- ChatGPT and Claude are both considered cloud models (with the exception of their downloaded versions). But these aren’t the only cloud models available for use with Vibe coding. Google’s Gemini models can also help recognize and write code. They support VSCode only through Antigravity (Google’s IDE similar to VSCode).
- Use RunPod and other GPU rental providers. Enterprising individuals recognized early on that not everyone could afford to purchase machines with the appropriate GPU power to run models effectively. Services like RunPod were created to fill this gap. Ollama can run on RunPod, and you can configure your VSCode to use those models. The main downside to this approach is that the models on Ollama are not as robust as the frontier models (OpenAI, Anthropic, Google, etc.). But I have tried them for Vibe coding, and they work okay.
Ollama was mentioned earlier as a local model, but they recently announced they also have a cloud version. You would need to pay for inference on this cloud version, but it gives you yet another option should you run out of credits on the other options.
It is said that timing is everything in life, which is also true of Vibe coding models. If you work with several models, they often have daily limits and weekly limits. If you haven’t used one (like Codex, for example) and you have a lot of credits left before the weekly reset, and the others have several days before their reset, it’s time to use Codex (or whichever one will reset its credits soon). If you are not Vibe coding full-time, try to schedule most of your updates for the day before the model resets.
If you are Vibe coding full-time, you probably don’t have this luxury of timing, as you need to use your credits when you need them. Your company probably already has a higher allocation of credits across multiple providers for you.
Should You Pay for Extra Credits?
Like most people, you likely have several subscriptions to various AI software solutions. You could have a ChatGPT subscription, Claude, Gemini, Midjourney, and a host of others. These subscriptions add up to serious money for each individual. So, it’s a hard pill to swallow when we’re told that if we run out of credits to buy more.
I get it and I was a stickler for not paying. You may be, too. But at least the option is available. You may come across a situation where you’re knee-deep in a coding session with the AI, and then you run out of time. That’s disruptive to the creative process and will likely derail the motivation for that session.
Having that extra little push for usage will avoid that. Of course, don’t wait until the extra usage is up to put yourself into the situation again. Use the extra usage time to finish any outstanding workflows, then have the AI update your docs so you can start up again when the weekly allocation resets (or use another model).
Your Experiences May Be Different
You’ll likely find your own groove when it comes to your Vibe coding experiences. What I present here are suggestions I have used, and they have helped. But they may not be right for your situation, or maybe some are, and the others aren’t, and that’s okay. All these experiences are still fairly new to all of us. Others will likely publish what works for them, and their ideas may not be similar to mine, which is good. The point is, though, there are ways to get around the new changes, even if all the players jump on board to implement them.
Conclusion
I briefly mentioned running local models with software like Ollama. If you don’t have enough horsepower, i.e., a decent GPU, on your local machine, RunPod is a great alternative. You can choose from a variety of GPU configurations, and Ollama can be installed directly from templates as you set up your machine. You don’t pay for credits, but you do usually rent machines by the hour (you can get a decent 3090 machine that will handle most of your Vibe coding needs, for 69 cents per hour (USD)).
