My AI Budget: The Before and After

A few months ago, I kept running out of session tokens: the amount of AI use my plan allows in a few hours of work. I upgraded my plan, and that helped. Then I started running out of weekly tokens instead.

My plan has two limits. One covers each five-hour session, and the other covers the whole week, as the plan's own help page explains. When I upgraded, I could do much more in each session, but every session still counted against my weekly limit. It felt like a contradiction. You would think you could max out your sessions and still stay under your weekly limit, but that is not how it works.

So in October I made a run of changes, and I wrote about them in Putting My AI on a Budget Without Touching the Checks. Three of them matter most here.

1. Dynamic Modeling. Before a task is handed out, the AI judges how hard it is and picks a model to match. Small, mechanical work goes to a cheaper model, and hard work stays on the strongest one. The checks always stay on the strongest model. It started Oct 12.
2. Chat limits. Each reply re-sends the whole chat so far, so a long chat costs more on every message. A monitor now warns me when a chat gets too long, and the AI writes a hand-off note so a fresh chat can pick up where the old one stopped. It started Oct 16.
3. RAG. It stands for retrieval-augmented generation: the AI searches my written plans and decisions first, then answers from what it found. I added it mainly so the AI would stop acting on a ruling that was later overturned, which a plain hunt through my records could turn up. By my October post, it had not yet paid back what it cost to build. It also started Oct 16.

Here is the before and after. I did this while working on real projects, so it is not a controlled test with perfect answers. It is real-world use, with some numbers to show what changed.

How to Read the Numbers

• These figures cover every session my main AI coding tool logged on my laptop. Work done with other tools is not in these counts.
• Every dollar figure is what the same use would cost at pay-per-use list prices. I pay a flat subscription, so none of it is money I spent or saved.
• A prompt is not a unit of work. This counts how often I typed, not how much got done.
• Hand-typed Prompts are every prompt I typed. Work Prompts leave out the lines that only start my day or carry work into a fresh chat after a hand-off, because the chat limits create more of those. Their cost still counts. The logs can only guess a hand-off from its timing, so Work Prompts are sometimes a range.
• After is Oct 16 to Nov 1, once all three changes were on.
• Before comes in three versions, because the changes did not all start on the same day:
  • Before Dynamic Modeling: Oct 1 to 11.
  • Before Chat Limits + RAG: Oct 1 to 15. It includes the first four days of Dynamic Modeling.
  • First week only: Oct 1 to 6. It matches no change. I show it because the early days were uneven, and the starting point moves the answer.
• None of these is a clean before. A setting that keeps the AI's replies short was already on from around Oct 6. The oldest logs have also partly aged out, so early-October totals read low, and an early price per prompt could be off in either direction.

Each Prompt Got Cheaper, and I Typed More

What this means. Each prompt got cheaper. A Work Prompt cost $10.91 at list price before Dynamic Modeling, and $8.04 to $8.29 after all three changes were on. Counting every Hand-typed Prompt, it went from $10.04 to $6.39. Depending on which Before I use, a Work Prompt cost about 14% to about 27% less, and a Hand-typed Prompt about 30% to about 38% less.

I also typed more. By my logs, Work Prompts per day worked rose about 6% to about 35%, depending on the Before, and Hand-typed Prompts rose about 30% to about 57%. So a working day cost about the same: about $300 at list price against Before Dynamic Modeling, and a little less against the other two. I also worked more days, so my use across the calendar went up.

Less Old Chat in Every Prompt

What this means. A token is a small chunk of text. Each reply re-sends the whole chat so far, plus the rules and instructions the AI started with. Most of that is read back from a stored copy, and my logs count it as a re-read. Before the changes, about 95% of all tokens were re-reads. After, about 93% were. Re-reads are billed at a small fraction of the price of new input. On the strongest model I use, it is one-twentieth, by the maker's published price list. That is why the dollar chart above matters more than the token count.

Each figure in this chart is all the tokens my logs recorded in that period, agents' work included, divided by the Work Prompts I typed. So it is an average per prompt, not the size of one request. All tokens per Work Prompt fell from 22.6 million to 13.8 to 14.2 million, about 26% to about 39% fewer, depending on the Before. Fresh tokens, the new text the AI takes in, fell much less: about 2% to about 20%. So most of the drop is less old chat being re-read in every prompt. That is the part the chat limits target, though other changes landed at the same time. By that measure, my prompts got leaner: each one carries less old chat.

Dynamic Modeling shows up in the model mix. Cheaper models went from about 1% of AI calls to about 14%, but they carry only about 6% of the cost. The strongest model still does most of the expensive work, and the checks stay on it.

The Week It Went Back Up

What this means. The drop was not smooth. The cheapest week was Oct 19 to 25, at $5.33 to $5.59 per Work Prompt. Then Oct 26 to Nov 1 went back up to $10.35 to $10.43, close to where I started.

I wanted to know why the price went back up, so I had the logs dug into. The short answer is: partly long agent jobs. My median prompt, the middle one if you line them all up by time, still took under ten minutes from when I typed it to its last step, in both weeks. But prompts that ran over an hour went from about 6% of my prompts to about 12%, and from about 36% of the cost to about 65%. Almost every one of them was a prompt where the AI sent the work out to my Agent Team.

I had quite a few long jobs that took days. The logs see prompts, not jobs, and a job that runs for days shows up as many shorter prompts. So they can't tell whether those multi-day jobs drove the rise, and "partly" is as far as they go.

Did It Help My Weekly Limit?

This part is from memory, not from a record. Before the changes, I hit my weekly limit at least once in October. As of November 3, 2026, I have not hit it since. I still come close, since a lot of my work runs overnight, done by my Agent Team. It seems to take longer to get there, though, and the meter seems to climb no faster per day, maybe a little slower, on more use. I also worked more days. My usage logs can't see that meter, so they can't prove or disprove this, and a change on the vendor's side would look the same.

What the logs can show is my list-price use. A week of it is the price per prompt times the number of prompts, and I sent more prompts. So the lower price per prompt did not shrink my week on its own. It paid for the extra prompts. At list price, my use per calendar day went up about 3% to about 29%, depending on the Before.

One More Number

At list price, my October use would have cost $8,773.34, about 44 times my $200-a-month plan. That is only what my logs still held. The oldest October logs have partly aged out, so both numbers are the least the month could have been. On pay-per-use pricing, a bill like that adds up fast. I would guess most people doing this much work are on a subscription.

What This Means for Your Team

To me, the changes were worth it.

Adding Dynamic Modeling, RAG and Chat limits has been worth the effort and tokens used to put this in place. Several changes landed together, and the mix of work may have changed too, so no one change gets the credit, and I can't put a number on RAG by itself. These numbers come from my own setup. They are not a forecast for your team.

If you track your team's AI use, look at cost per day beside cost per prompt. A lower price per prompt can hide rising use.

I am looking to package my technology for wider use. If you are interested, reach out to me and let me know.

If You Want to Talk

If you have a project that is stuck, or work you have won and can't staff, I would like to hear about it.

I start with a two-week assessment of one codebase and how work moves through it, including how your team is using AI-written code. You get a written findings document with a prioritized plan.

Send me a message on LinkedIn, or book a 20-minute call.