How Should a Growing Company Budget for AI Tools: API, Subscription, or Local Models?

Subscription plans from the model makers are currently the cheapest way to buy AI capacity, because the subsidy sits inside the price. Direct API billing costs the most. Open-source models you host yourself cost the least per token and the most in engineering time. Treat the choice as a decision you revisit quarterly.

The harder part of the budget is that the arithmetic changes when a vendor merges, loses model access, or reprices, and you find out after your team has already built its working habits on top of it.

Why AI Costs Can Blow Through a Budget Fast

AI spend behaves differently from the software line items you're used to. A seat-based tool costs what it costs whether your team logs in or not. An agent costs more the better your team gets at using it, because competence shows up as volume: more sessions, longer sessions, larger context windows, more parallel runs.

Stephen Sutzer, lead operator at oAT, has watched this land on real budgets. As he puts it: "Uber blew through their entire like AI budget in the in like Q1 for the entire they blew their through their entire budget for the year in Q1."

Three mechanics drive the overrun:

• Loops, not calls. An agent working through a task reads files, plans, executes, checks its work, and retries. One human request can be hundreds of model calls.
• Adoption compounds. The pilot group's usage in month one is a poor predictor of month four, once the rest of the team sees the output.
• Frontier models get used for everything. Teams reach for the most capable model available, including for work a cheaper one would finish.

API vs. Subscription: What's Actually Subsidized

Direct API billing looks like the honest price because it is metered per token [1]. In practice it is the ceiling. Subscription products from the same companies bundle a usage allowance that the provider is willing to sell below cost while it competes for adoption.

Stephen is blunt about the gap: "your subscriptions are essentially subsidized. like your your token costs are essentially subsidized by the model makers up to I think some estimates are like 7,000% that they're subsidizing if you're if you're billing directly via the API it's always going to be the most costly."

What follows from that for your budget:

• Buy subscription capacity first for any workload that fits inside a plan's limits [2].
• Reserve API billing for what genuinely needs it: production features, automated pipelines, anything running without a human in the seat.
• Expect the subsidy to shrink. Price a scenario where your current usage costs meaningfully more, and know which workloads you'd cut.
• Track usage per person and per workflow, so a repricing event is an adjustment rather than a surprise.

The Three Billing Models, Side by Side

The Vendor Lock-In Risk: What Happened With Cursor

Price is one exposure. Continuity is the other, and it moves faster. Stephen describes an engagement where a whole organization had standardized on one coding tool:

"one organization that we've we've worked with went all in on Cursor. Uh, and then whenever Cursor and uh, SpaceX AAI essentially merged and you know, SpaceX XAI is has bought Cursor and is incorporating that into their their platform. OpenAI released that like starting in November, they were no longer going to be servicing cursor uh, their their agent harness with OpenAI models."

The lesson is about layers. When you adopt an AI tool you are adopting at least three things that can move independently: the interface your team has learned, the agent harness that runs the loop, and the underlying model. A commercial change at any layer can strand the other two.

Questions worth asking before you standardize:

• If the model behind this tool became unavailable next quarter, what would we do on Monday?
• Are our prompts, workflows and evaluation criteria stored somewhere portable, or only inside the vendor?
• How much of our team's skill is transferable, and how much is muscle memory for one interface?
• Who at our company owns noticing that a vendor relationship has changed?

Should You Build Your Own Stack?

When owning more of the stack pays

Self-hosted open models earn their keep on repetitive, high-volume, well-defined work: classification, extraction, summarization inside a known document set, anything where the task barely changes and the quality bar is measurable [3]. If a workflow runs thousands of times a month on the same shape of input, you can serve it locally and stop paying frontier prices for a task that doesn't need frontier reasoning.

What it costs you elsewhere

The bill moves from the vendor invoice to your payroll and your hardware. Someone has to select, serve, evaluate and update models, and keep doing it as better ones ship [4]. For most growing companies that headcount is the real price, and it competes with the work that actually differentiates the business.

The practical middle

Most teams land on a mix: subscriptions for human-in-the-loop work, API for production paths, local models for the narrow high-volume jobs, and a written rule for which workloads sit where. The mix matters less than having someone who owns it and revisits it.

Map the Process Before You Price the Tools

The fastest way to overspend is to roll out agents into a process nobody has mapped. Stephen's framing from the field is that deploying without a plan just accelerates the mess: you get more words, more code and more documentation than before, without more value.

At oAT we approach client work from the systems level first, asking what problem is actually being solved and what steps the work takes, then deciding where agents fit, where automation works, and where humans stay in the loop [5]. That mapping is also your budget model. Once you know how many times a month a workflow runs and how many model calls each run takes, you can price it. Until then you're guessing.

Where Your People Sit Changes What You Spend

Stephen points to Anthropic's own analysis of how people use coding agents effectively: the humans who get the most out of these tools spend their time in planning and review, while the agent does the bulk of the execution [6]. In his account of that research, humans spent roughly 70% of their time planning and the agent spent roughly 80% of its time executing.

That split has a budget consequence. The moment someone on your team starts building or using agents, they effectively move up a layer in the org chart, and strategy, judgment and taste become the job. Spending on planning and review capability, which means people and their time, is what keeps the token spend productive. Skip it and you pay twice: once for the generated output, once for the cleanup.

how We Do: Budgeting AI Tooling in Four Steps

1. Inventory the workflows, not the tools. List the processes you want agents inside. For each one, note how often it runs and who reviews the result. This is the unit you will price.
2. Assign each workflow a billing model. Human-in-the-loop work goes to subscription capacity. Unattended production paths go to metered API with hard caps. Narrow high-volume jobs are candidates for self-hosting.
3. Set a portability floor. Keep prompts, process maps and evaluation criteria in your own systems. Adopt any single interface only as deep as you can afford to lose it.
4. Re-price every quarter. Put a recurring 60-minute review on someone's calendar with a named owner. Usage data, vendor changes and new model releases all land inside a quarter.

Signals That It's Time to Re-Price

FAQ

Why did Uber blow through its entire AI budget in one quarter?

The internal details are not public. The pattern oAT's Stephen Sutzer describes is that agent-driven usage scales with adoption: as more of a team learns to work with agents, each request turns into many model calls, and a forecast built on pilot-stage usage understates the real number by a wide margin.

Are AI subscriptions actually cheaper than paying per API call?

Currently, yes. According to oAT's Stephen Sutzer, subscription token costs are subsidized by the model makers, by some estimates up to 7,000%, while billing directly through an API is always the most costly route. The sensible read is that the subsidy is a competitive tactic and will not last forever, so any budget should model what the same usage costs without it.

What happened when Cursor lost access to OpenAI's models?

As oAT's Stephen Sutzer describes it, one organization had gone all in on Cursor, and then Cursor was acquired by xAI and folded into that platform, while OpenAI announced it would stop servicing Cursor's agent harness with OpenAI models starting in November. Teams that had standardized entirely on one tool had to change their working setup on a vendor's timetable rather than their own.

Should a growing company commit to a single AI vendor or model provider?

Standardize on one tool for a given workflow, because shared habits and shared prompts make teams faster. Keep the underlying assets portable: process maps, prompts and evaluation criteria belong in systems you own. That combination gives a team the speed of a single standard with the option to move when a vendor relationship changes.

References

1. OpenAI. "API Pricing." https://openai.com/api/pricing/
2. Anthropic. "Pricing." https://www.anthropic.com/pricing
3. Meta. "Llama." https://www.llama.com/
4. Ollama. "Ollama." https://ollama.com/
5. of All Trades. "Services." https://weofalltrades.com/services.html
6. Anthropic. "Anthropic Economic Index." https://www.anthropic.com/economic-index