More Context Isn't Better: You Pay for Extra Twice

Why does one AI task cost $10 over three weeks, while another costs $100 in a single hour? Same tool. Same person using it. The difference was the context.

High AI costs are often just a symptom. The real problem is poor context management.

Everyone says to give AI as much context as possible. But more context isn’t always better. Two of my own projects taught me exactly why.

Three Weeks, Ten Dollars

I was researching how to use hardware keys from within a Docker container - with Claude, in a single chat. It sounds simple. It wasn’t. There were a lot of problems to work through. That research ran three weeks and around 150 iterations, all in that one session - until the moment I went to make some last fixes, and Claude refused, because…

I didn’t even know that was possible. 😅 And yeah, I still use the web chatbot for some things.

Resulting price: about $10.

This also challenged a piece of common advice - keep one chat per task, because long conversations supposedly make models worse. But it’s not always that black and white.

A lot depends on the use case. In my story, there was one target the whole time. The more I progressed and the more I tested, the narrower the scope became - and the better my outputs got.

Length has real limits - I hit one, as you saw. But within those limits, the thing that degrades output isn’t length itself, it’s the conversation drifting from the original goal. A focused conversation ages far better than a wandering one.

And if you feel the AI drifting away from the real problem, instead of steering it, you can ask it to write a memo. It can pick that up at the start of a new session - you get everything necessary, without the surrounding conversation that no longer adds value.

But I also have a very different story - one that cost 10x more, and I had no idea why.

One Hour, One Hundred Dollars

The first story piled up context for weeks and stayed cheap. This one was four files that barely changed - and it blew up.

Size was never the variable.

I needed to verify that the old version of the FE behaved the same way the new one does - or to find the differences. We’d migrated some APIs that required FE changes because of a data-structure change. I reached for Playwright - a tool I’d never used before, but it fit perfectly.

The idea was simple. The code was simple too.

But this became the biggest outlier I’ve ever had in token consumption and cost. In about an hour it went from 10 thousand tokens to 100 million. How is that even possible?

Sometimes what matters isn’t the project you’re using AI on - it’s the way you use it.

On the surface, all the AI did was run one script - npx playwright test. What I didn’t watch was what it did in between:

Nice, Test 56 looks OK, I will keep monitoring.

I let it handle everything - starting the script, reading the output, analyzing it. And it didn’t wait for the script to finish. It resent the whole context on every cycle, and since Playwright is loud, each cycle carried a mountain of output.

Instead, I could have run the script myself and asked the AI to compare the final results. That would have sent one complete output instead of dozens of partial ones.

This is a great example of huge - but worthless - context.

Your context shouldn’t hold the most information it can. It should hold only what that one task needs. That can be hard to define, and even harder to provide.

A bloated context costs you twice: money you can see, and clarity you can’t.

Watch What You Can’t See

AI context is like currency, and managing it is a new skill we need to acquire. Normally you’d never spend money without knowing what you’re spending it on - but with AI, that somehow became normal.

And it’s not just about money. A context stuffed with useless data has less room for what actually matters - which is how you start getting worse answers, not just bigger bills. So watch your consumption, yes - but watch your context too.

The difference between the $10 and the $100 wasn’t the output. Both times I got useful answers - what changed was the context I fed in. The $100 story could have been a $10 one.

Using AI is like writing. The value was never in the words themselves - it’s in how they’re arranged. Same goes for context.