How to Maximize Your Coding Agent Subscriptions

A common issue I started facing when using a coding agent such as Claude Code or Codex was that my usage was going extremely fast. By extremely fast, I mean that for Codex, I would typically use up my weekly allowance in about 10-12 hours, which becomes very expensive if I am to cover everything for a week with my Codex subscriptions.

I started thinking about what causes my subscriptions to go that fast, and I think there are numerous reasons for this, for example:

• My code repository gets larger over time.

• We get access to better models, such as Claude Fable or GPT-6-Astra, which cost more to run and thus increase the usage faster.

• The companies are less generous with the subscription coding plans, as they cost them a lot of money to run.

• We simply code more, and more agents in parallel.

Thus, I had to figure out a way to make my coding and subscription plans last longer while impacting quality minimally.

This article will be a summary of my experiences with coding agents and how I get more out of my subscriptions, and the specific techniques that I apply every day when I code.

Why do coding agent subscriptions empty so fast now?

First, let’s cover why this is an issue at all. I recall the beginning of the year, in the period between January and April at least, as being very lenient with the usage limits. I would have one Claude account and one Codex account, and they would get me through the week without a problem. I didn’t really manage to hit the usage limit on the 20x subscriptions on either of them within a week.

Now, however, I’ve regularly hit the usage limits, and as I mentioned earlier in the article regarding Codex weekly usage, I have been able to hit it in 10 hours just the other day, and I have the same experience when I use the Claude subscription; I am definitely able to use it up within one day.

If my subscriptions only last one day each, that would require me to have seven subscriptions, which would of course cost quite a lot of money to maintain per month, considering they each cost $200 USD per month.

I think there are many reasons why coding agent subscriptions empty faster now, and I think part of it is that the companies are being less generous with the subscriptions now that we have access to smarter models, such as Claude Fable and GPT-6 Astra. However, I also think that we as programmers have started to program even more. We’re producing more code, running more agents in parallel, and thus emptying our usage faster.

Now, of course, I don’t want to reduce the amount of coding that I do, because the whole point of having these subscriptions is to be more productive. So, I need to figure out other ways of getting more out of the same subscriptions without impacting quality, which is what I’ll discuss in the next section.

How to code more with the same subscriptions

In this section, I’ll discuss the specific techniques that I apply to get more out of my coding agent subscriptions. On a high level, I use the following techniques:

• Use smaller models for simpler tasks (GPT-5.6 SOL, Claude Opus)

• Use the smartest models (Fable and GPT-6) for orchestration, and make sure they spin up subagents with smaller models

• Refactor the codebase regularly (complicated codebase, requires more reasoning from the models, costing more tokens, etc.)

• Avoid god files

• Clean up your markdown files (AGENTS.md, CLAUDE.md, skills, hooks, etc)

Now let’s dive deeper into each of these. So first of all, I would like to comment on the fact that when I say the smartest models, the frontier models, I refer to the two models, Claude Fable and GPT-6-Astra, which, at the moment of writing this article, are by far the two best models available on the market. In my opinion, there’s nothing even close to the capabilities of these two models. But these are also the models that cost the most in terms of usage.

When to use smaller and when to use larger models

First, let’s cover when you should use smaller models and when you should use larger models. I think you should always try to use smaller models whenever you’re doing simpler tasks. For example, if you’re just researching something on the web, doing a simple computer-use task, or exploring a topic, I think you should definitely try to use the smaller models. When I say smaller models, I would refer to GPT-5.6 SOL or Claude Opus, which are still very capable, but not on the level of Fable and GPT-6 Astra. It’s worth noting that I don’t personally use models that are even smaller than that, such as GPT Luna, GPT Terra, and Claude Sonnet, because I just don’t believe they’re capable enough for the tasks that I want to do.

If I want to do more complicated tasks, I start spinning up Fable and GPT-6. However, it’s very important to note that I don’t always use Fable and GPT-6 for all coding within that complicated task; I always, when trying to spin up these models, try to tell them to use sub-agents that are smaller models. So if I’m using Claude Code, I tell it explicitly to spin up Opus sub-agents whenever it does the work, and Fable just acts as the orchestrator. The same goes for GPT-6 Astra: I don’t fully use the model for all tasks; when interacting with the task, I ask it to spin up GPT-5.6 SOL sub-agents to research topics, write specific code implementations, and so on. This saves an enormous amount of tokens and allows you to get way more out of your subscriptions. In my opinion, it also doesn’t really impact quality that much compared to the orchestrator model. It’s most important that it’s a smart model because that model can command smaller models and give them very specific tasks to implement, which the smaller models are still very good at. It’s not strictly necessary to use the smartest models to do such small, specific, targeted tasks.

To start doing this yourself, you can tell your Claude and Codex right now that whenever they spin up sub-agents, they should always spin up Opus or GPT-5.6 SOL as sub-agents and not Fable or GPT-6 sub-agents. Just doing this quick little change will vastly increase the amount of usage you get out of your subscriptions.

Optimizing your codebase

The second main tip that I urge you to start implementing is to optimize your codebase so that agents can work as effectively as possible. Now, this is not only about the amount of usage you get out of your subscription, but it’s also about how easy it is for your agent to perform implementations in your codebase. So, I have three specific recommendations here.

1. Regularly refactor your codebase (once a week at least)

2. Clean up god files (refactoring)

3. Clean up MD files

Number 1 is that you should refactor your codebase at least once a week. So I suggest that you sit down every weekend with your agent, tell it to look at the codebase, look at the refactoring opportunities and stuff you can do to optimize your codebase to make development better, easier, and faster. You can find a bunch of prompts online on how to refactor your codebase effectively, and just running these prompts, telling the agent to implement them, driving it autonomously to dev, and verifying that the code works the same before and after, will help you a lot in getting more usage out of your coding agents and making it easier for them to work in your codebase. So this will also have the additional benefit of making your coding agents faster at implementing code in your codebase.

One particular refactor that I do very regularly is to look for god files, i.e., files that are very big and contain a lot of code. This typically takes a lot of usage for three reasons. First of all, whenever the agent reads this large file, it costs a lot of tokens to read this file, and thus it increases the usage for the model. Secondly, god files have a tendency to be complicated, so it will be harder for the agent to work there. And thirdly, god files will be changed often simply because a lot of code and logic are in those files. This leads to a lot of agents doing overlapping work within these files and having to rebase on top of each other, which again costs a lot of usage. So, due to these three reasons, god files should be avoided at all costs, and you should very regularly clean them up, look for large files, split them up, and thus save a lot of usage.

Luckily, it’s very easy to look for refactoring opportunities and solve them because you can simply tell your coding agents to do it. However, you do have to perform the explicit action of asking your coding agents to do the refactoring as they likely won’t do that themselves.

The third tip I gave you is to clean up your markdown files. When I say clean up your markdown files, I mean clean up CLAUDE.md, AGENTS.md, and clean up your skills if you don’t use them anymore. Look to see if you have any hooks that are not being used anymore or that you don’t want to use anymore. Claude has a command you can run, which is:

This command just runs a specific prompt that cleans up all the skills and markdown files that are not being used anymore. You can also just find a standardized prompt online to clean this up.

Conclusion

In this article, I’ve discussed how to get more out of your coding agent subscriptions such as Claude Code and Codex. I started experiencing, maybe a few weeks to a few months ago, that I was getting less and less out of my coding agent subscriptions and had to take immediate action to get more out of my subscriptions, more usage without sacrificing quality. I thus started implementing specific techniques to optimize the amount of coding I got out of my coding agents, which are the techniques that are discussed here today. For example, using smarter models as orchestrator agents and smaller models to perform the grunt work, such as implementing code or researching specific topics. I also urge you to refactor your codebase regularly to make sure agents can run effectively in your codebase. If you implement the techniques that I covered in this article, you’ll definitely notice the amount of usage you get out of your subscriptions goes up, which makes coding agents cost less per month.

👋 Get in Touch

👉 My free eBook and Webinar:

🚀 10x Your Engineering with LLMs (Free 3-Day Email Course)

📚 Get my free Vision Language Models ebook

💻 My webinar on Vision Language Models

👉 Find me on socials:

💌 Substack

🔗 LinkedIn

🐦 X / Twitter