For at least some codebases, I'm not sure this is a useful metric. Because you don't usually put the whole codebase in your context at the same time.
For example in my current case, there are lots of files with CSS, SVG icons in separate files, old database migration scripts, etc. Those don't go in the LLM context 99% of the time.
Maybe a more useful metric would be "what percentage of files that have been edited in the last {n} days fit in the context"?
On a related note, this type of reasoning is what made me flip my opinion on microservices. I've generally been skeptical of a many-microservice architecture for the last decade but LLMs change that - a small microservice is more likely to fit in a context window.
I think this gestures at a more general point - we're still focusing on how to integrate LLMs into existing dev tooling paradigms. We squeeze LLMs into IDEs for human dev ergonomics but we should start thinking about LLM dev ergonomics - what idioms and design patterns make software development easiest for AIs?
Microservices are about deployment, less about code structure. You can have the same code modularization as microservices provide within a monolith instead, for example in the form of libraries.
Maybe it’s useful to dig out the concept of modularization with a distinction between interface and implementation again, and construct agents that are able to make effective use of it.
In the case that interfaces remain unchanged, agents only need to look at the implementation of a single module at a time plus the interfaces it consumes and implements. And when changing interfaces, agents only need to look at the interfaces of the modules concerned, and at most a limited number of implementation considerations.
It’s the very reason why we humans invented modularization: so that we don’t have to hold the complete codebase in our heads (“context windows”) in order to reason about it and make changes to it in a robust and well-grounded way.
Maybe it could just measure the number of tokens for the examples (and then summarize what the examples show, under the assumption that that’s the actual functionality of the project). I’m 90% joking… but that last 10% makes me wonder…
Useful and useless (or good and “less good”) aren’t easily mapped to big and small.
From a purely UX perspective, showing a red badge seems you’re conflating “less good” with size. Who is the target for this? Lots of useful codebases are large.
I do agree, however, that there’s value in splitting up domains into something a human can easily learn and keep in their head after, say, a few days of being deeply entrenched. Tokens could actually be a good proxy for this.
Outside of packages I doubt few of my code bases would fit into this. But the individual domain areas would. I don't care about users in a orders context, I don't care about payments when dealing with imports, no reason an ai should care either. It shouldn't care about implementations if there's an interface referenced, it shouldn't worry about front end when it's dealing with the back etc.
Scoping the Ai to only use the things you'd use seems far wiser than trying to reduce your codebase so it can look at the whole thing when 90% of it is irrelevant.
This is a really interesting metric to track. I agree with the sentiment that token budgets are becoming the new 'lines of code' metric. Even though context windows are constantly expanding (like the 200k default you used for Opus), there's still a tangible benefit to keeping a codebase lean. It's not just about fitting it into the window, but also about the signal-to-noise ratio for the agent. The color-coding based on percentage is a nice touch for a quick visual health check.
It’s interesting but I think it’s measuring the wrong thing. Abstraction is a fundamental principle in software. As a human, I’ve worked with classes and modules far larger than what fits in my head, just because I’m only fitting the function signatures and purpose into my head, and not the implementation details. In practice I find Claude really good at extracting useful information in a human-like way from a codebase. It doesn’t usually stuff the entire codebase into its context window.
Also this rewards dynamic languages over typed languages, penalizes comments, descriptive function names, etc. Though frankly, it'd be interesting to see whether AI would work better with a project in Javascript that barely fits in context, or the same thing in typescript that overflows. I could imagine either, but my guess is "it depends". Though, "depends on what" would be interesting to know.
Still, this seems useful for being able to see at a glance. I have no idea where most of my own projects would land.
I haven't cared too much about repo tokens in a good while.
But my coolest app was a better context creator. I found it hard to extend to actual agentic coding use. Agentic discovery is generally useful and reliable - the overhead of tokens can be managed by the harness (i.e. Claude Code).
This is an interesting concept. Thank you for sharing. I have an export.sh or export.ps1 script that takes the relevant files in my repository and puts them in a `dump.txt` file inside `docs/llm`.
I am not very good with AI though. Is there a quick and easy way to calculate token count and add this to my dump.txt file, ideally using just simple, included by default Linux tools in bash or simple, included by default Windows tools in powershell?
Gemini 1.5 Announced the 1 million token context window in 2024. I admire this view of being forward looking towards new technologies, specially when we see the history of how bad people can be at predictions just by looking at history HN posts/comments.
If we look at back 2 years, companies weren't investing into training their LLMs so heavily on code. Any code they got their hands on was what was in the LLMs training corpus, it's well known that the most recent improvements in LLM productivity occurred after they spent millions on different labs to produce more coding datasets for them.
So while LLMs have gotten a lot better at not needing the entire codebase in context at once, because their weights are already so well tuned to development environments they can better infer and index things as needed. However, I fail to see how the context window limitation would no longer be an issue since it's a fundamental part of the real world. Would we get better and more efficient ways of splitting and indexing context windows? Surely. Will that reduce our fear of soiling our contexts with bad prompt response cycles? Probably not...
I’m not so sure an increasingly large context window will be seen as a critical enabler (as it was viewed 6 months ago), after watching how amazingly effective subagents and tool calls are at tackling parts of the problem and surfacing the just the relevant bits for the task at hand. And if increasing the context window isn’t the current bottleneck, effort will be put elsewhere.
I agree. My suspicion is that token efficiency is what will drive more efficient tool calls, and tool building. And we want that. Agents should rely less on raw intelligence (ability to hold everyting in context), and more on building tools to get the job done.
Some say that the ideal size of an individual function in a codebase is related to the amount of information you can hold in working memory. Maybe the ideal size for a library is the amount you can fit in an LLM context window?
> What’s the going rate for tokens in terms of dollars?
It depends on the provider/model, usually pricing is calculated as $/million tokens with input/output tokens having different per token pricing (output tends to be more expensive than input). Some models also charge more per token if the context size is above a threshold. Cached operations may also reduce the price per token.
The math on what people are actually paying is hard to evaluate. Ime, most companies rather buy a subscription than give their developers API keys (as it makes spending predictable).
> Ime, most companies rather buy a subscription than give their developers API keys (as it makes spending predictable).
The downside with subscriptions is that your work with the LLM will grind to a halt for a number of hours if you hit the token limit. I was doing what I consider very trivial work adding Javadoc comments to a few dozen files using Claude Sonnet on the $20 plan and within 30 minutes had been told to sit out for a couple hours. The reason was that Claude was apparently repeatedly sending the files up and down to fill in the comments. In hindsight, sure, that's obvious, but you would think that Claude would be smart enough to do some sort of summarization to make things more efficient. Looking into it, it was on the order of several million tokens in a very short amount of time.
It really made me wonder how in the hell people are using Claude to do "real" work, but I've heard of people having multiple $200/month subscriptions, so I guess that could work. Definitely seems like a glimpse into the future of what these services will truly cost once people are hooked on them.
For at least some codebases, I'm not sure this is a useful metric. Because you don't usually put the whole codebase in your context at the same time.
For example in my current case, there are lots of files with CSS, SVG icons in separate files, old database migration scripts, etc. Those don't go in the LLM context 99% of the time.
Maybe a more useful metric would be "what percentage of files that have been edited in the last {n} days fit in the context"?
On a related note, this type of reasoning is what made me flip my opinion on microservices. I've generally been skeptical of a many-microservice architecture for the last decade but LLMs change that - a small microservice is more likely to fit in a context window.
I think this gestures at a more general point - we're still focusing on how to integrate LLMs into existing dev tooling paradigms. We squeeze LLMs into IDEs for human dev ergonomics but we should start thinking about LLM dev ergonomics - what idioms and design patterns make software development easiest for AIs?
Microservices are about deployment, less about code structure. You can have the same code modularization as microservices provide within a monolith instead, for example in the form of libraries.
Maybe it’s useful to dig out the concept of modularization with a distinction between interface and implementation again, and construct agents that are able to make effective use of it.
In the case that interfaces remain unchanged, agents only need to look at the implementation of a single module at a time plus the interfaces it consumes and implements. And when changing interfaces, agents only need to look at the interfaces of the modules concerned, and at most a limited number of implementation considerations.
It’s the very reason why we humans invented modularization: so that we don’t have to hold the complete codebase in our heads (“context windows”) in order to reason about it and make changes to it in a robust and well-grounded way.
Maybe it could just measure the number of tokens for the examples (and then summarize what the examples show, under the assumption that that’s the actual functionality of the project). I’m 90% joking… but that last 10% makes me wonder…
Blogged about modular code and LLMs few days ago
https://www.slater.dev/2026/02/relieve-your-context-anxiety-...
Useful and useless (or good and “less good”) aren’t easily mapped to big and small.
From a purely UX perspective, showing a red badge seems you’re conflating “less good” with size. Who is the target for this? Lots of useful codebases are large.
I do agree, however, that there’s value in splitting up domains into something a human can easily learn and keep in their head after, say, a few days of being deeply entrenched. Tokens could actually be a good proxy for this.
> Who is the target for this?
Agents. Going to be more tools and software targeted for consumption by agents
Yeah, but a large monorepo can consist of many small subprojects. And arguably this is becoming a best practice.
Just spawn the agent in one of the subprojects
Outside of packages I doubt few of my code bases would fit into this. But the individual domain areas would. I don't care about users in a orders context, I don't care about payments when dealing with imports, no reason an ai should care either. It shouldn't care about implementations if there's an interface referenced, it shouldn't worry about front end when it's dealing with the back etc.
Scoping the Ai to only use the things you'd use seems far wiser than trying to reduce your codebase so it can look at the whole thing when 90% of it is irrelevant.
Im curious if there is a deep need for entire codebase to be consumed in the first place?
It would be better to have the architecture support a more decoupled/modular design if you're going to rely heavy on LLMs.
That or let it consume high quality maintained documentation?
This is a really interesting metric to track. I agree with the sentiment that token budgets are becoming the new 'lines of code' metric. Even though context windows are constantly expanding (like the 200k default you used for Opus), there's still a tangible benefit to keeping a codebase lean. It's not just about fitting it into the window, but also about the signal-to-noise ratio for the agent. The color-coding based on percentage is a nice touch for a quick visual health check.
It’s interesting but I think it’s measuring the wrong thing. Abstraction is a fundamental principle in software. As a human, I’ve worked with classes and modules far larger than what fits in my head, just because I’m only fitting the function signatures and purpose into my head, and not the implementation details. In practice I find Claude really good at extracting useful information in a human-like way from a codebase. It doesn’t usually stuff the entire codebase into its context window.
Also this rewards dynamic languages over typed languages, penalizes comments, descriptive function names, etc. Though frankly, it'd be interesting to see whether AI would work better with a project in Javascript that barely fits in context, or the same thing in typescript that overflows. I could imagine either, but my guess is "it depends". Though, "depends on what" would be interesting to know.
Still, this seems useful for being able to see at a glance. I have no idea where most of my own projects would land.
I haven't cared too much about repo tokens in a good while.
But my coolest app was a better context creator. I found it hard to extend to actual agentic coding use. Agentic discovery is generally useful and reliable - the overhead of tokens can be managed by the harness (i.e. Claude Code).
https://prompttower.com/
This is an interesting concept. Thank you for sharing. I have an export.sh or export.ps1 script that takes the relevant files in my repository and puts them in a `dump.txt` file inside `docs/llm`.
I am not very good with AI though. Is there a quick and easy way to calculate token count and add this to my dump.txt file, ideally using just simple, included by default Linux tools in bash or simple, included by default Windows tools in powershell?
Thank you in advance.
It's a fun, in the "style of the time" thing to track, but within a year or two, context window limitations won't be a thing.
Doubt me?
Think back 2 years. Now compare today. Change is at massive speed, and this issue is top line to be resolved in some fashion.
Gemini 1.5 Announced the 1 million token context window in 2024. I admire this view of being forward looking towards new technologies, specially when we see the history of how bad people can be at predictions just by looking at history HN posts/comments.
If we look at back 2 years, companies weren't investing into training their LLMs so heavily on code. Any code they got their hands on was what was in the LLMs training corpus, it's well known that the most recent improvements in LLM productivity occurred after they spent millions on different labs to produce more coding datasets for them.
So while LLMs have gotten a lot better at not needing the entire codebase in context at once, because their weights are already so well tuned to development environments they can better infer and index things as needed. However, I fail to see how the context window limitation would no longer be an issue since it's a fundamental part of the real world. Would we get better and more efficient ways of splitting and indexing context windows? Surely. Will that reduce our fear of soiling our contexts with bad prompt response cycles? Probably not...
I’m not so sure an increasingly large context window will be seen as a critical enabler (as it was viewed 6 months ago), after watching how amazingly effective subagents and tool calls are at tackling parts of the problem and surfacing the just the relevant bits for the task at hand. And if increasing the context window isn’t the current bottleneck, effort will be put elsewhere.
I agree. My suspicion is that token efficiency is what will drive more efficient tool calls, and tool building. And we want that. Agents should rely less on raw intelligence (ability to hold everyting in context), and more on building tools to get the job done.
Some say that the ideal size of an individual function in a codebase is related to the amount of information you can hold in working memory. Maybe the ideal size for a library is the amount you can fit in an LLM context window?
I'm not sure that smaller bases are always better.
value/size
What’s the going rate for tokens in terms of dollars? How much are companies spending on “tokens”?
Also kind of ironic that small codebases are now in vogue, just when google monolithic repos were so popular.
> What’s the going rate for tokens in terms of dollars?
It depends on the provider/model, usually pricing is calculated as $/million tokens with input/output tokens having different per token pricing (output tends to be more expensive than input). Some models also charge more per token if the context size is above a threshold. Cached operations may also reduce the price per token.
OpenRouter has a good overview over provider and models, https://openrouter.ai/models
The math on what people are actually paying is hard to evaluate. Ime, most companies rather buy a subscription than give their developers API keys (as it makes spending predictable).
> Ime, most companies rather buy a subscription than give their developers API keys (as it makes spending predictable).
The downside with subscriptions is that your work with the LLM will grind to a halt for a number of hours if you hit the token limit. I was doing what I consider very trivial work adding Javadoc comments to a few dozen files using Claude Sonnet on the $20 plan and within 30 minutes had been told to sit out for a couple hours. The reason was that Claude was apparently repeatedly sending the files up and down to fill in the comments. In hindsight, sure, that's obvious, but you would think that Claude would be smart enough to do some sort of summarization to make things more efficient. Looking into it, it was on the order of several million tokens in a very short amount of time.
It really made me wonder how in the hell people are using Claude to do "real" work, but I've heard of people having multiple $200/month subscriptions, so I guess that could work. Definitely seems like a glimpse into the future of what these services will truly cost once people are hooked on them.
Api keys with hard limits I assume?
Are there companies out there that add token counts to ticket “costs”, i.e. are story points being replaced/augmented by token counts?
Or even worse, an exchange rate of story points to tokens used…
Interesting concept, but is it going to age well with context sizes of models are changing all the time (growing, mostly)?
max context sizes are probably going to go up, but smaller contexts will always be cheaper/more-efficient than larger ones
Interesting, but not adding something to my CI for a badge, too paranoid.
Smart idea. Token budgets are becoming the new line count metric for the LLM era.
Nah. I can write a whole program using 0 tokens, I can’t write a whole program with 0 lines of code.