A standard problem I began going through when utilizing a coding agent equivalent to Claude Code or Codex was that my utilization was going extraordinarily quick. By extraordinarily quick, I imply that for Codex, I might usually expend my weekly allowance in about 10-12 hours, which turns into very costly if I’m to cowl every little thing for every week with my Codex subscriptions.
I began interested by what causes my subscriptions to go that quick, and I believe there are quite a few causes for this, for instance:
-
My code repository will get bigger over time.
-
We get entry to raised fashions, equivalent to Claude Fable or GPT-6-Astra, which price extra to run and thus improve the utilization sooner.
-
The businesses are much less beneficiant with the subscription coding plans, as they price them some huge cash to run.
-
We merely code extra, and extra brokers in parallel.
Thus, I had to determine a approach to make my coding and subscription plans last more whereas impacting high quality minimally.
This text will likely be a abstract of my experiences with coding brokers and the way I get extra out of my subscriptions, and the particular methods that I apply day-after-day once I code.
This infographic highlights the principle contents of this text. I am going to talk about the way to get extra out of your coding agent subscriptions and the particular methods that I apply each day to maximise my coding agent effectivity. Picture by ChatGPT
Why do coding agent subscriptions empty so quick now?
First, let’s cowl why this is a matter in any respect. I recall the start of the yr, within the interval between January and April at the least, as being very lenient with the utilization limits. I might have one Claude account and one Codex account, and they’d get me by means of the week with no drawback. I did not actually handle to hit the utilization restrict on the 20x subscriptions on both of them inside every week.
Now, nevertheless, I’ve recurrently hit the utilization limits, and as I discussed earlier within the article concerning Codex weekly utilization, I’ve been in a position to hit it in 10 hours simply the opposite day, and I’ve the identical expertise once I use the Claude subscription; I’m undoubtedly ready to make use of it up inside at some point.
If my subscriptions solely final at some point every, that may require me to have seven subscriptions, which might after all price fairly some huge cash to keep up per thirty days, contemplating they every price $200 USD per thirty days.
I believe there are lots of the reason why coding agent subscriptions empty sooner now, and I believe a part of it’s that the businesses are being much less beneficiant with the subscriptions now that we’ve got entry to smarter fashions, equivalent to Claude Fable and GPT-6 Astra. Nevertheless, I additionally suppose that we as programmers have began to program much more. We’re producing extra code, working extra brokers in parallel, and thus emptying our utilization sooner.
Now, after all, I do not wish to scale back the quantity of coding that I do, as a result of the entire level of getting these subscriptions is to be extra productive. So, I would like to determine different methods of getting extra out of the identical subscriptions with out impacting high quality, which is what I am going to talk about within the subsequent part.
The best way to code extra with the identical subscriptions
On this part, I am going to talk about the particular methods that I apply to get extra out of my coding agent subscriptions. On a excessive degree, I exploit the next methods:
-
Use smaller fashions for less complicated duties (GPT-5.6 SOL, Claude Opus)
-
Use the neatest fashions (Fable and GPT-6) for orchestration, and ensure they spin up subagents with smaller fashions
-
Refactor the codebase recurrently (sophisticated codebase, requires extra reasoning from the fashions, costing extra tokens, and so forth.)
-
Keep away from god information
-
Clear up your markdown information (AGENTS.md, CLAUDE.md, abilities, hooks, and so forth)
Now let’s dive deeper into every of those. So to begin with, I wish to touch upon the truth that once I say the neatest fashions, the frontier fashions, I check with the 2 fashions, Claude Fable and GPT-6-Astra, which, in the mean time of writing this text, are by far the 2 greatest fashions accessible in the marketplace. In my view, there’s nothing even near the capabilities of those two fashions. However these are additionally the fashions that price essentially the most by way of utilization.
When to make use of smaller and when to make use of bigger fashions
First, let’s cowl when you must use smaller fashions and when you must use bigger fashions. I believe you must at all times attempt to use smaller fashions everytime you’re doing less complicated duties. For instance, in case you’re simply researching one thing on the net, doing a easy computer-use job, or exploring a subject, I believe you must undoubtedly attempt to use the smaller fashions. After I say smaller fashions, I might check with GPT-5.6 SOL or Claude Opus, that are nonetheless very succesful, however not on the extent of Fable and GPT-6 Astra. It is price noting that I do not personally use fashions which might be even smaller than that, equivalent to GPT Luna, GPT Terra, and Claude Sonnet, as a result of I simply do not imagine they’re succesful sufficient for the duties that I wish to do.
If I wish to do extra sophisticated duties, I begin spinning up Fable and GPT-6. Nevertheless, it is crucial to notice that I do not at all times use Fable and GPT-6 for all coding inside that sophisticated job; I at all times, when attempting to spin up these fashions, attempt to inform them to make use of sub-agents which might be smaller fashions. So if I am utilizing Claude Code, I inform it explicitly to spin up Opus sub-agents at any time when it does the work, and Fable simply acts because the orchestrator. The identical goes for GPT-6 Astra: I do not absolutely use the mannequin for all duties; when interacting with the duty, I ask it to spin up GPT-5.6 SOL sub-agents to analysis matters, write particular code implementations, and so forth. This protects an unlimited quantity of tokens and lets you get far more out of your subscriptions. In my view, it additionally would not actually influence high quality that a lot in comparison with the orchestrator mannequin. It is most vital that it is a sensible mannequin as a result of that mannequin can command smaller fashions and provides them very particular duties to implement, which the smaller fashions are nonetheless excellent at. It isn’t strictly essential to make use of the neatest fashions to do such small, particular, focused duties.
To start out doing this your self, you may inform your Claude and Codex proper now that at any time when they spin up sub-agents, they need to at all times spin up Opus or GPT-5.6 SOL as sub-agents and never Fable or GPT-6 sub-agents. Simply doing this fast little change will vastly improve the quantity of utilization you get out of your subscriptions.
Optimizing your codebase
The second foremost tip that I urge you to begin implementing is to optimize your codebase in order that brokers can work as successfully as doable. Now, this isn’t solely concerning the quantity of utilization you get out of your subscription, nevertheless it’s additionally about how straightforward it’s in your agent to carry out implementations in your codebase. So, I’ve three particular suggestions right here.
-
Frequently refactor your codebase (as soon as every week at the least)
-
Clear up god information (refactoring)
-
Clear up MD information
#1 is that you must refactor your codebase at the least as soon as every week. So I counsel that you simply sit down each weekend along with your agent, inform it to have a look at the codebase, have a look at the refactoring alternatives and stuff you are able to do to optimize your codebase to make improvement higher, simpler, and sooner. You could find a bunch of prompts on-line on the way to refactor your codebase successfully, and simply working these prompts, telling the agent to implement them, driving it autonomously to dev, and verifying that the code works the identical earlier than and after, will assist you numerous in getting extra utilization out of your coding brokers and making it simpler for them to work in your codebase. So this can even have the extra profit of creating your coding brokers sooner at implementing code in your codebase.
One specific refactor that I do very recurrently is to search for god information, i.e., information which might be very massive and comprise quite a lot of code. This usually takes quite a lot of utilization for 3 causes. Initially, at any time when the agent reads this massive file, it prices quite a lot of tokens to learn this file, and thus it will increase the utilization for the mannequin. Secondly, god information tend to be sophisticated, so will probably be tougher for the agent to work there. And thirdly, god information will likely be modified typically just because quite a lot of code and logic are in these information. This results in quite a lot of brokers doing overlapping work inside these information and having to rebase on high of one another, which once more prices quite a lot of utilization. So, because of these three causes, god information needs to be prevented in any respect prices, and you must very recurrently clear them up, search for massive information, break up them up, and thus save quite a lot of utilization.
Fortunately, it’s totally straightforward to search for refactoring alternatives and resolve them as a result of you may merely inform your coding brokers to do it. Nevertheless, you do should carry out the specific motion of asking your coding brokers to do the refactoring as they possible will not do this themselves.
The third tip I gave you is to scrub up your markdown information. After I say clear up your markdown information, I imply clear up CLAUDE.md, AGENTS.md, and clear up your abilities in case you do not use them anymore. Look to see when you have any hooks that aren’t getting used anymore or that you do not wish to use anymore. Claude has a command you may run, which is:
This command simply runs a particular immediate that cleans up all the talents and markdown information that aren’t getting used anymore. You may also simply discover a standardized immediate on-line to scrub this up.
Conclusion
On this article, I’ve mentioned the way to get extra out of your coding agent subscriptions equivalent to Claude Code and Codex. I began experiencing, perhaps a couple of weeks to a couple months in the past, that I used to be getting much less and fewer out of my coding agent subscriptions and needed to take rapid motion to get extra out of my subscriptions, extra utilization with out sacrificing high quality. I thus began implementing particular methods to optimize the quantity of coding I received out of my coding brokers, that are the methods which might be mentioned right here at present. For instance, utilizing smarter fashions as orchestrator brokers and smaller fashions to carry out the grunt work, equivalent to implementing code or researching particular matters. I additionally urge you to refactor your codebase recurrently to verify brokers can run successfully in your codebase. When you implement the methods that I coated on this article, you will undoubtedly discover the quantity of utilization you get out of your subscriptions goes up, which makes coding brokers price much less per thirty days.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day E mail Course)
📚 Get my free Imaginative and prescient Language Fashions e-book
💻 My webinar on Imaginative and prescient Language Fashions
👉 Discover me on socials:
💌 Substack
