Putting My AI on a Budget Without Touching the Checks

My AI kept running up against its weekly allowance before the week was over, and at least once it ran out. So I changed how I use it. In my main chats, the history the AI re-reads on an average call fell from about 390,000 tokens to about 147,000 (October 1–11 against October 16–20).

On October 16, I put the problem plainly: "I am constantly eating up my weekly quota early and need to curb that." The checks on the AI's work were the one thing I would not cut.

I use Claude Code with Anthropic's models.

I pay a flat subscription, so this was about room, not money. Every dollar figure here is what the same use would cost at pay-per-use list prices. None of it is money I spent or saved.

Long Chats Cost More Than They Look

A long chat costs more on every message, even a one-line one. About 95% of all tokens were the AI re-reading what it had already read: the chat so far, and the rules and instructions it started with.

Re-reading is billed at a small fraction of the price of new input. Storing each new part of the chat so it can be re-read later makes up about half of the estimated cost.

On October 16, I rebuilt a chat-length monitor. At 150,000 tokens, the reply opens with a red box: finish this step, then start a new session. At 250,000, the AI is set to write a handoff note first: where it stopped, what is saved, the exact next step, and any open decisions. Then the red box tells me to open a new session and paste one line pointing to the note.

The warning is red on purpose. My screen is full of logs that I like to watch, and a plain warning would get lost in them.

Before October 16, 12 of 30 chats peaked above 400,000 tokens. Since then, none of 101 has (through October 22), which shows the limit is being followed.

Match the Model to the Task

On October 12, I added dynamic model routing. The coordinating agent judges how hard each task is and picks a model. Small, mechanical work, like finding files, goes to a cheaper model. Routine work goes to a middle tier. Hard work stays on the strongest model: architecture, security, compliance, and money.

The checks are the exception. My rules keep code review, the test gate, and the security and compliance check on the strongest model, however small the change. Cheaper models can build. They do not decide whether the work passes. That keeps the reviews where they were; it does not prove they catch as much as before.

Calls on cheaper models rose from about 1% of all calls (October 1–11) to about 7% (October 16–20). The first days cost more, not less: about $299 per day worked at list price (October 1–11), then about $425 (October 12–15), while I built the changes and my chats grew.

For mechanical steps, my rules now turn the model's thinking effort down before reaching for a cheaper model. I added that setting on October 21.

My projects run with AI agents, and one agent coordinates them. While building model routing, I accidentally let my main chat start the agents itself, so every full report landed in my main chat and made it longer. Since October 21, the coordinating agent starts the agents and reads their reports. I get one plain-English summary with the action items.

What the Checks Caught

Three of the builds went through a separate code review, and all three were sent back for fixes on the first round.

One review found a gap. Each agent has a written list of the tools it may use, but the way my agents were being started, those lists were not enforced: every agent could use every tool. That review was of a change meant to reduce usage, which is why I kept the reviews.

Now every brief repeats the agent's allowed tools, and the agent names the tools it used on the first line of its answer. A tool outside the list is a defect, and the task is re-run. This is a check, not a lock: an agent could still use the wrong tool, and the check relies on the agent's own report, after the fact.

The checks catch ordinary delivery mistakes too. On October 24, the separate code review of a change to my site's menu found that, on older browsers, the menu would have stayed open over the page with buttons that did nothing. It was fixed before it shipped.

The Rulebook and the Search

My written rulebook loads at the start of every chat. Every rule I add is read every time, like a subscription. By an estimate from my usage logs, it made up about 30% of the text loaded when an agent started.

It had filled up with history: where each rule came from, and what went wrong before it. On October 21, I moved that history into hidden notes in the same file, which the tool strips out before loading. Every rule still loads, none was cut or weakened, and the loaded text shrank 16.5% in size. A separate review checked the result, and a fresh load showed none of the hidden text came through.

On October 16, I switched on a RAG layer over my written plans, decisions, and lessons: the AI searches the written record first, then answers from what it finds. A result is a pointer, never proof: the AI must open the file before it acts, and a decision that has since been replaced comes back marked as old.

I use it daily, but its accuracy has not been formally tested. On October 31, I plan to check it against 20 known questions, including replaced decisions, with a script scoring the answers, not an AI.

What It Took, and What It Returned

Around October 6, I turned on a setting for shorter replies, and the average reply fell from about 641 tokens to about 300. That is the least certain estimate below.

Update, November 3, 2026: this held from Oct 7 to 22. From Oct 23 the average across all AI calls rose to about 983 tokens, and my logs do not show why. The return estimates on this page run only to Oct 22.

These are my own estimates from my usage logs, at pay-per-use list prices, from each change's start to October 22.

Building all of it took about 275–700 million tokens, the baseline included: $132–334 at list price. My own time was about four hours, by my estimate, mostly keeping an eye on things.

By my own estimates, these changes avoided $413–704 in usage at list prices through October 22, so by then they had covered their estimated build cost. Some of those estimates are low-confidence. That compares usage with usage. It leaves out my four hours, and the extra rule text I added earlier in the month, about $55 a week at list price. Counting that rule text, the worst case had not quite covered its cost by October 22. Whether the work paid off for the business is not proven yet.

For the full method, the numbers by period, and the caveats, read the extended version.

What I Still Need to Measure

An average AI call got cheaper, from about 11 cents to about 8 at list price (October 1–11 against October 16–20). But a call is not a finished task, and one task can take many calls. Four things are left to measure:

• Delivery: finished, accepted work, set against usage.
• Quality: corrections, re-runs, the effect of lower effort, and anything lost at a handoff.
• My time: setup, stepping in, review, and upkeep.
• Capacity: how much useful work fits before the weekly allowance runs out.

Several changes landed in the same week, so no single change gets the credit, and one long chat holds 22% of October's tokens through October 22. My usage logs can't see the plan's own meter. Since the changes, I have not hit the cap, but I still come close by the end of each week.

On October 23, I checked: that day's work had used 7% of my weekly limit. By that time on a normal day, I am usually at about 14% and rising. That is one reading on one day, not a trend.

None of this changed what I decide myself: every release, and any push the scan flags.

What This Means for Your Team

These numbers come from my own setup. They are not a forecast for your team. None of these techniques is new, either. The work was picking the ones that fit, and checking that they did what I thought.

If your team uses AI every day, it is already making a trade, planned or not: which work gets the top model, and which work gets checked. Before you cut AI costs, find the steps that decide whether code ships, and leave those alone. Put the cheaper models on the legwork, where a mistake is cheap to catch.

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.