Token Maxing vs. Budget Maxing: How Leaders Should Think About AI Agent Costs

When OpenClaw first hit and everyone got access to capable models and harnesses, turning everything up to maximum was actually the right move.

Nobody really knew what these systems could do yet, so you wanted the most intelligence, the most tool use, and the deepest thinking available. That exploration phase, where a small group of practitioners and solopreneurs with some money to spend just played with everything turned up to ten, is what people started calling token maxing.

That approach does not scale once adoption spreads across a real business. I have been watching the signals for the past two months, and they all point the same way.

Tesla capped employee AI spending at $200 a week, with sign-off required to go above that. If you are already paying for a Claude Max plan and a top-tier OpenAI license, that is roughly $400 of the $200, so it is not a stingy number, but it is a hard ceiling where there was none before.

Uber had set a budget for the year and burned through the entire thing by April, months ahead of schedule, once agent tools and token maxing took hold. They had to reset the budget and put real caps in place.

Priceline started putting AI usage dashboards directly into its financial reviews so technology leaders could see what was actually being spent. Qualcomm adopted a showback model, where teams see the dollar cost of their own AI usage even if the money is not literally charged back to their budget.

None of that is really about whether people are using lots of tokens. It comes down to whether what they are doing makes the work better, for the employee and for the workflow, not just faster.

There is a difference between optimization AI, using AI to do the same work faster and cheaper, and opportunity AI, using the time you free up to rebuild the work itself. Both matter, but they are not the same thing.

Model routing is where a lot of this gets solved technically. Not every task needs the top foundation model, and for years we were trained to reach for whatever model just launched, the same way everyone jumped on GPT-5 or the newest Claude the day it shipped.

Factory, an AI coding platform, built a routing engine that sends each task to the most cost-effective model that can actually do the job. On Terminal-Bench 2, routing cut cost by roughly 20 percent while holding pass rate close to 99 percent of what the frontier model alone would score, a number nobody had put up on that benchmark before.

Cursor took a similar bet a step further. It trained its own coding model on top of an open Chinese model, Qwen 2.5 Coder, and now runs that model inside its own infrastructure at close to frontier quality, at a fraction of the cost and speed of calling out to a foundation model every time.

Here is the personal habit I actually use, and the one I teach in our courses once someone is AI fluent enough to be designing workflows and agents. Do not optimize while you are still building. Keep the intelligence and thinking turned up while you figure out the workflow itself and get it working end to end.

Once it works and I am happy with it, I ask the model or the agent harness directly what level of model and reasoning the workflow actually needs to run reliably. I test that recommendation. Almost every time, for things like reading email, checking a calendar, or pulling together a simple brief, the answer comes back as a small model on low reasoning effort, not Opus, not Fable.

That pattern holds across most of the workflow, not just the final model choice. Here is roughly how I split it once a workflow is up and running:

• Design and problem definition benefit from a strong reasoning model, because that is where you are talking through the approach.
• Gathering information and connecting tools does not need that same strength, so a smaller model on low reasoning handles it fine.
• Organizing files and structuring a system is well within reach of a small model too.
• Drafting is often better with a smaller model than people assume, since Haiku and Sonnet-class models write well without needing Fable or Opus.
• Review is the one place I keep a stronger model in the loop, because judgment is what you are actually paying for at that stage.

Good workflow design is the budget strategy. If you build it this way from the start, especially for anything that runs on a schedule, almost everything I run on a routine cadence ends up on a lower-tier model by default, and I only bump it up when there is a specific reason it needs more.

One of our members raised a sharp point during the session that is worth repeating here: a lot of token maxing is really people skipping the thinking step and going straight to building.

The old engineering habit of planning before you touch a keyboard has not gone away just because AI can generate output instantly. If anything, it matters more now, because an agent will happily burn tokens executing a plan you never actually thought through.

At Lead with AI, we are going through the same internal transformation, and it starts with people feeling safe enough to experiment before a policy tells them exactly what to do.

‍Brandon Sammut and Tracy St.Dic from Zapier and Chase Warrington at Doist both walked our PRO members through versions of this, and one detail from the Zapier session has stuck with me: strong guidance, not restrictive policy, is what gives people the psychological safety to actually build.

That structure gets captured as golden paths, the best workflows, tested and documented, that get released back to the whole team instead of five people separately solving the same problem in their own silo.

Layered on top of that is a context layer, a shared company brain containing strategy, priorities, policy, and the golden paths themselves, so nobody has to reassemble that context from scratch every time they start a new conversation or spin up a new agent.

Zapier even builds an AI-native readme for every employee, describing what that person does and is good at, so the company's own AI can be asked to find the right working group for a project by searching across everyone's readme. It is a small idea with a big multiplier once your workforce reaches real AI fluency.

Token maxing was the right instinct during exploration. It gets people building, experimenting, and discovering what agents can actually do. But at some stage, that has to convert into ROI, into repeatable skills, repeatable agents, and repeatable workflows that can be deployed responsibly across the business, not into a leaderboard of who used the most tokens this week.

• Build the workflow first with the model turned all the way up, and only optimize the model and reasoning level once the workflow is actually working.
• Ask your agent harness directly what level of model and thinking a task needs, then test that recommendation before you trust it.
• Reserve frontier models like Opus and Fable for design and review, and route routine, tool-heavy tasks to smaller, cheaper models by default.
• Measure the outcome your tokens produced, not the number of tokens spent, before you reward or celebrate heavy AI usage.
• Build a shared company context layer and a documented set of golden paths so teams stop rebuilding the same automation in isolated silos.

Gary Tate is Chief AI Officer at Lead with AI, where he leads product and internal AI transformation and delivers the monthly Agentic Briefing for Lead with AI PRO members.