Match the Model to the Job: Where AI Agent Tokens Go

Agent work spends tokens in three places, and each one can run on a different model at a different price. A leader paying for that work has to see all three to know where the money and the plan allowance go. Below is how I split them, with spend I measured, and what the split did to my bill and to my failures.

I'm Bob Michaels. I started building for the web in 1992, and today I rebuild how companies are understood on the web and by AI: content strategy, web architecture, AI-visibility research and custom agent systems. Across the many client projects I run at once, AI agents do much of the work under my direction. An agent is a model given a task and the tools to carry it out. Every figure comes from my own runs, in the system I built to manage them, and was current on 2026-10-15.

Where do the tokens in AI agent work go?

AI agent work spends tokens on three jobs: reading web pages and files to find facts, doing the work with what was found, and writing the report back at the end. A token is a small piece of a word. Vendors bill by the tokens a model reads and writes, and subscription plans meter use the same way. On a pay-per-token account, waste shows up on the invoice. On a plan, it shows up as an allowance that runs out sooner.

The report back is the easiest of the three to miss. It is text the agent writes, so the vendor counts it as output, the same as the work it describes. On a plan, every agent in a session draws on the same allowance, so one heavy job can leave the others short.

One total says how much was spent. To find where it went, split it by the three jobs first and then by vendor.

Which model should do the research in an agent system?

In my system, the smallest model does the research. I settled that on 2026-08-18, and it has been a standing decision since.

HALO treats each kind of job as a named type, such as writing, judging, research, planning or building, and I assign each type its model from a single control screen. The choice follows the kind of job, whichever project it sits in. Research turns up across projects, so setting it project by project would mean revisiting each one whenever I changed my mind. Setting it by type means the August decision covers all of them. It also keeps any one project from slipping onto a pricier research model.

One research run of 90 calls shows why spend has to be broken out by vendor. Once it finished, actual spend by vendor came to OpenAI $2.66, Grok $1.31, Anthropic $1.03, Perplexity $0.50 and Gemini $0.28, about $5.78 in total. The total alone would not have told me where to cut. The Anthropic model was not the expensive vendor that time. The money went to input, the text OpenAI and Grok were handed to read, so shrinking that input volume is where cutting cost starts.

What happened when expensive AI agents did their own research?

On 2026-09-23, the agents used up my whole session allowance partway through the pass. I had sent twenty expensive writer agents out at once. Each one did its own web research and read a 45-page source in full. Every one of them paid for its own reading at the expensive rate, and all of it came out of one allowance. Ten of the agents died in flight: they stopped before they finished.

My system has a guard that checks which model an agent will use before it goes out. It passed every dispatch, every agent it was asked to send, correctly, because the guard was checking the model and not the plan. A model check and an allowance check answer different questions, and I only had the first.

The fix was to split the job. Cheap agents find and verify the facts and return the source plus the verbatim sentence, the exact words in the source that state the fact. The expensive agent writes from what they returned, so the writer does no looking up of its own. A fact that comes back with its source and exact sentence can be checked against that source by anyone who opens it.

Each wave is capped at six to eight agents, a limit I set after that day, and I wait for it to finish before sending the next. If your agents run on a subscription plan, ask whether anything checks the allowance before a batch goes out, separately from any check on which model each agent uses.

Why does the report back from an AI agent cost tokens of its own?

An agent's reply is output, and the vendor bills it even when it repeats work already saved to a file. A 52-agent compliance round, run the same day as the writer pass, showed how fast that adds up. Each agent wrote its verdict to its own file in full, then wrote the same reasoning again in its reply.

Several replies ran past 4,000 words, and the system truncated, or cut short, most of them. The full copy was already on disk, so every verdict was paid for twice, and most of the second copies arrived incomplete. A reply costs tokens separately from the work. The rule now is that the file is the result and the reply is three lines. If your agents save their work to files, look at how long their replies run.

Raw data follows the same rule. I save the raw results of every run to disk. When I re-score a run or re-analyze it, the work reads the saved file instead of paying a vendor again for data I already bought.

How do you know what an AI agent run really cost?

Read what each vendor actually billed once the run is over, and treat any figure of zero as suspect. That billed spend is what I report after every run, not the estimate.

A zero once fooled my own system. A model with no row in my price table, the list my system uses to turn token counts into dollars, reported a cost of zero on every call while it was billing at $3 per million input tokens and $15 per million output tokens. I wrote up how that unpriced model read as free. Any call that reads and writes text uses tokens, so a zero means a price is missing somewhere. A reading of zero dollars is now treated as a bug until proven otherwise.

A tier is a size and price class of model. Before any fan-out, a batch of agents sent out at once, I ask three questions: which tier does the looking up, which tier does the work, and what does the report cost. Your team can ask the same three before its next batch goes out and write the answers down. The answers are short: a tier for each of the first two, and for the third, whether the reply repeats what the agent already wrote down. If the most expensive tier is doing its own reading, change that first. After the run, hold those notes against the vendor bills.

Common questions