The Best Way to Reduce Your Company's AI Token Spend

The best way to reduce your company’s AI token spend is to build a custom model you own. This allows companies to maintain control over cost, access, and intellectual property.

AI tokens are data units, including words, characters, and punctuation marks, that large language models read and produce. Models bill input tokens (what you send) and output tokens (what the model returns) at separate rates. Ramp, which processes AI vendor payments for thousands of businesses, reports that output costs three to five times as much as input.

Company AI bills keep growing because token usage is rising faster than per-token prices are falling. Among businesses with connected AI accounts, Ramp found that token usage grew 1,001 percent from January 2025 to April 2026, while total spend grew 497 percent.

On top of that, users have shifted from basic prompting to full-on agentic workflows that complete multi-step tasks, choosing their own tools and number of steps. This multiplies the tokens each task uses. In a Financial Times article on strategies to deal with AI costs, Adaption co-founder Sara Hooker said, “Agentic workflows are much more compute-intensive and much more unpredictable about the amount of compute they need.”

AI token prices will not fall without action from companies. And if they do decrease, they won’t fall reliably. Per-token prices have moved in both directions over the past 18 months, and company bills have kept growing either way.

Hooker expects token usage to keep growing and prices to rise. “There’s a clear understanding in the industry that tokens are not priced properly right now. Once the IPOs [of frontier labs like OpenAI and Anthropic] happen, there is going to be immediate pressure on those margins to be more sensible.”

Two mechanisms can reduce token spend: adjusting how you use a provider’s models or building your own custom model. Adjustments can lower the number of tokens you use or the price you pay per token, but they’re not predictable. Only building a custom model gives you total control over AI spending.

A custom AI model is a model trained on a company’s proprietary data to do the work your company actually does, such as extracting clauses from contracts, classifying customer requests, or processing records in bulk. It replaces a general-purpose model with one built for your business.

Custom models are different from general-purpose models because they are specific to your context. Fine-tuning, training an existing model on new data, is one common way to build one. They help companies avoid applying one very large model to every problem.

Speaking at Fortune Brainstorm Tech in June 2026, Hooker said that “probably 90% of problems are very easy,” and she named bulk processing as a common example where a massive model is not needed.

A custom model can cost less to run because a model built for routine work can be much smaller than a general-purpose model, and its price no longer depends on one provider’s price list.

In September 2026, the research team at Adaption, which builds AI systems that adapt and self-improve through real-time interaction, tested how far a smaller model can go through harness optimization. On the Harvey Legal Agent Benchmark, the open model Qwen3.6-27B, which has 27 billion parameters, passed 67 percent of criteria across 140 held-out tasks. After harness changes and post-training on top, it passed 88 percent, close to frontier performance.

This points to a simple conclusion: with the right setup and training, a smaller model can come close to frontier performance on the work it is built for, so model size alone does not decide quality.

Move a workload to a custom model when it is routine work, the volume is steady and high, you have training examples, and you can test quality. Keep a workload on a frontier model when it needs broad reasoning that a smaller model fails on, runs rarely, or changes every week.

“Probably 90% of problems are very easy. Many things that you do in bulk processing, for example, you shouldn’t be throwing a massive model at,” Hooker shared with Fortune in June 2026. The useful question is not which workloads can leave a frontier model, but which ones need it.

Four questions to consider:

Adaption makes it fast and easy to build a custom, self-improving AI model. Building one takes three things: a defined task, training examples, and a training method. Adaption’s products cover the data and training steps, so you do not need a research team.

Prepare Your Training Examples with Adaptive Data or Invent a Dataset

Building a custom model means you must have AI-ready training data. Adaption offers two products for this. Adaptive Data optimizes, enhances, and expands raw data that a company already has. It has shown to deliver 82% quality improvement across workflows on average. If you have no existing data, Invent a Dataset builds a training dataset from a description of the behavior you want, with no seed data, predefined schema, or labeling guide needed.

Adjust your training recipe with AutoScientist

AutoScientist is Adaption's automated model training system. State the behavior you want, and the system adjusts the training data and training settings together, run after run, until the model meets that objective. AutoScientist takes a developer from idea to an owned, adapted model in an afternoon. In a preliminary experiment for AI Singapore on a small base model, AutoScientist saved 56 engineering hours of manual training-parameter optimization.

Training a model used to be hard to get right outside a frontier lab. Automated data preparation and training now handle much of that work. What is left for your team is the part only you can do: choosing the task, supplying or approving examples, and testing the result against your own requests. If a workload is not a fit for a custom model, the next section covers how to reduce spend with the models you already use.

A custom model is not the right move for every workload. For those cases, you can reduce spend by changing how you use the models you already pay for.

Token Spend Tracking

Token spend tracking means recording how many tokens each workflow, team, and model uses, and what they cost. Tracking itself does not lower spend. It allows you to see which workflows drive the bill and decide where a change would matter most.

Token spend tracking takes setup: usage has to be labeled by workflow and team across every provider and API key. Workflows that share a key are hard to separate. Additionally, provider reports show usage after the fact, so a runaway agent can run up a bill before anyone sees it.

Trimming Prompts, Context, and Output

Trimming means removing content the model does not need from each request and limiting how much it writes back. Tokens hide in four places: the system prompt, retrieved documents, chat history, and the length of the answer. Trimming sends only what the task needs, asks for short structured output when software reads the result, and uses ordinary code instead of a model call for fixed tasks, such as reading a date from a field.

Trimming has limits. A shorter prompt that drops a needed instruction can lower answer quality, and the rework can cost more than the tokens saved. It also only reduces the cost of the content you cut. It does not change the price per token or which model you pay for. The work is ongoing too, because prompts and chat histories can grow back as teams add instructions and features.

Routing Tasks to Smaller Models

Model routing means sending each request to the cheapest model that can do the job well. A rule or a small program decides which model handles each task, so simple work such as classification or formatting goes to a smaller, cheaper model and hard work goes to a premium one.

Routing saves money only when the smaller model passes your quality tests, so it needs a test set of real requests and a pass mark for each task. A hard task sent to a weak model produces errors that can cost more to fix than the routing saves. The router is also one more system to run, and it needs re-testing when prices, models, or tasks change. You still pay the provider’s price per token on every request.

Caching Repeated Prompt Content

Prompt caching is a provider feature that stores the processed beginning of a prompt and charges less when a later request repeats it. Putting the system prompt, tool definitions, and reference documents first and the changing part last gives the cache something to reuse.

Caching helps only when many requests share a long, stable beginning, and workflows where every input is new get little benefit. The cache also expires: Anthropic’s default lifetime is five minutes, so requests that arrive far apart miss it. Caching lowers the price of repeated input only, not output, and you still depend on the provider’s pricing.

Batching Non-Urgent Work

Batch processing means submitting a group of requests together and getting the results later, at a lower price. Results can take up to 24 hours, so it’s best for work that can wait, such as nightly summaries or evaluation runs, and not live chat or anything a person is waiting on. Features can also be restricted: OpenAI’s Batch API does not support streaming, and OpenAI’s zero data retention policy does not apply to it, which matters for companies with strict data rules. The discount is a fixed percentage off the provider’s price, so the provider still sets the price.

Capping Agent Runs

Capping agent runs means setting a limit on the steps and tokens each run can use, and a spending limit for each team that runs agents. Because each step an agent takes is billed, and an agent chooses its own tools and number of steps, a cap stops a looping or wandering agent before it runs up a large bill.

Caps control the worst case, not the average cost, so an agent that stays under the cap can still waste tokens on every run. A cap set too low cuts off legitimate long tasks and returns unfinished work, and a cap set too high does little. Choosing a sensible cap requires knowing how many steps a normal run takes. Caps also do not address why agents waste steps.

For routine, high-volume work, building a custom model is the best way to reduce your company’s AI token spend, because it gives you complete control over your costs. Making adjustments – such as capping agent runs and trimming prompts, context, and output – can help with workloads that stay on a general model, but your company is still at the mercy of a provider’s pricing and terms. A custom model changes the terms themselves, and Adaption makes this path practical.

Talk to an AI expert at Adaption about moving your first workload.