Sam Altman said in June that customers have burned through their entire 2026 AI budget already. That has been consistent with my experience - token costs and economics has been a prime concern at every CTO level conversation I have had lately.
I have been talking about deterministic vs agentic AI for a while, largely with a focus on how it is more reliable, but the cost conversation really makes this shine too.
Deterministic vs Agents
An agent works by putting the reasoning power on the model. You hand it a goal and a prompt, and the model decides the next step from there, until it decides the goal is done. That flexibility is what makes it so valuable. The model can handle problems whose shape you have not fully defined up front, and in many cases can deliver. The trade-off is the reason this works. The model reasons at runtime on every turn, each tool call runs a large amount of tokens on top of that, and token usage climbs.
Deterministic AI is the other shape. We take the process apart and define it up front in code. We direct the process ourselves and give the model specific, highly bounded, atomic sub-tasks. This is where LLMs actually shine - it's the first time we can make a thinking function, and with a single targeted goal they're both cheap and successful. This means there is no decision-making at runtime and no variance in the path. The full breakdown of this distinction sits in an earlier blog.
So why does this help?
We can estimate the cost up front
Because the workload is a set of component parts and we know how each one runs and we can estimate the cost range up front. This gives us a known cost to work against, and we can run a cost-benefit analysis up front before major builds or deployments, comparing the outcome cost to what the process used to run at.
Given that over time our business starts to rely on these automations operationally, a clear forecast on cost variance lets us model the future in the most accurate way possible and scale as we need to.
The model is not baked in
We have written before about how new model versions change their reasoning, and how that pulls the ground out from under an agent leaning on a specific model's judgement. Those same model changes affect costing too, in two ways.
First, the models themselves get more expensive. Sonnet 5 launched in June at $2/$10 per million tokens. In September the standard rate lifts to $3/$15, a 50% jump on input - your workload is now 50% more expensive and all that extra burn will start to hurt.
Second, newer models tend to reason more aggressively, solving problems in more complex ways and burning more tokens per solve. That is fine on a deterministic line where a station only asks the model to complete a specific task, but on an agentic build the reasoning runs over the whole workload.
On a deterministic line you swap the one station that needed a frontier model onto Haiku 4.5, Mistral, or another cheaper model, and the rest of the line does not move. We are not locked into an escalating cost model the way an agentic build softly is.
We are not sacrificing the cool stuff
Doing this work is not sacrificing the cool stuff. It is what makes the cool stuff run. Part of our worldview is that agents need rails to run on, especially the more complex ones. Deterministic automations are those rails. They become an inventory of priced, trusted tools with known behaviour that the agent can lean on, so the agent focuses on the part where it shines, orchestration. That work is what we mean by L1.
An agent picks a tool and runs it. If the tools do not exist, the agent is guessing at runtime, and paying to guess. The agentic setups that work in production are the ones sitting on top of L1.