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

[Adaption Team](https://www.linkedin.com/company/adaption-labs/)

·

Oct 19, 2026

·

10min read

![The Best Way to Reduce Your Company's AI Token Spend](https://adaptionlabs.ai/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2F5fn3so8t%2Fproduction%2Fec77a7c17db1abe1905fc013b0e1a62f9288b79f-1780x866.jpg&w=1920&q=75)

The best way to reduce your company’s AI token spend is to build your own custom model. This lets companies control cost, access, and intellectual property while reducing vendor lock-in.

### Key Takeaways:

- Company AI bills keep growing: token usage grew 1,001 percent from January 2025 to April 2026, while total spend grew 497 percent.
- Adaption co-founder Sara Hooker says costs will keep rising because token costs are artificially low, and frontier labs will have to manage margin expectations after their upcoming IPOs.
- There are two ways to reduce your company’s AI token spend: 1) adjust how you use a provider’s model or 2) build a custom model you own. Only building a custom model allows you to permanently control costs.
- Building custom models is easier with Adaption, which makes data optimization, data generation, and automated model training seamless with its Adaptive Data, Invent A Dataset, and AutoScientist platforms.

### What Is an AI Token?

An AI token is a unit of text, such as a word, word fragment, or punctuation mark, that a large language model reads or writes, and providers bill by the token. Models charge separate rates for input tokens (what you send) and output tokens (what the model returns).

### Why Do Company AI Bills Keep Growing?

Company AI bills keep growing because token usage is rising faster than per-token prices are falling. Two factors push volume up:

- Usage volume: Among businesses with connected AI accounts, Ramp found token usage grew 1,001 percent from January 2025 to April 2026, while overall spend grew 497 percent.
- Agent steps: An agent completes multi-step tasks, chooses its own tools, and decides how many steps to take, and each step is billed. A [2026 Stanford University study](https://digitaleconomy.stanford.edu/publication/how-do-ai-agents-spend-your-money) of eight frontier models on coding tasks found agentic tasks consumed about 1,000 times more tokens than code chat.

In an October 2026 Financial Times article on[ strategies to ](https://www.ft.com/content/3effd703-591f-43c5-828e-dd11839653eb)manage AI costs, Sara Hooker, co-founder of Adaption, which builds continual learning systems, said agentic workflows need more compute than single prompts. As a result, costs are harder to predict.

### Will AI Token Prices Fall?

Effective token prices are falling now, but they have moved in both directions, and the provider sets them. [Ramp's AI Index](https://ramp.com/data/ai-index-sept-2026) shows the effective price businesses pay per million tokens fell 41 percent, from a 2026 peak of $1.15 in March to $0.68 in early September, after providers announced price cuts.

Two further things moved the price:

- Customer trade-down. Frontier models' share of tokens fell from a 53 percent peak in August to 45 percent as some businesses set defaults toward standard models.
- Launch-driven increases. [Fortune](https://fortune.com/2026/09/09/ai-compute-tokens-cheaper-boom/) reported that effective prices rose after major model launches this year before easing.

Hooker expects token usage to keep growing and prices to rise, in part because the planned IPOs of frontier labs will pressure their margins. She told the Financial Times, “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.”

A custom model's running cost follows its own size and hosting, so a provider's repricing in either direction does not change it. Falling prices do raise the monthly token spend a workload needs before a custom model repays its training and hosting costs.

### What Are Your Options for Reducing Token Spend?

There are two mechanisms for reducing token spend: adjust how you use a provider’s models or build your own custom model. Making adjustments lowers the number of tokens you use or the price you pay per token, but is not predictable. Only building a custom model gives you total control over AI spending.

| Option                            | What it changes                                                       | Works when                                                                          | Limits                                                                                 |
| --------------------------------- | --------------------------------------------------------------------- | ----------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------- |
| Build a custom model              | Which model runs the task; removes dependence from external providers | Routine work runs at steady volume and you have (or can generate) training examples | Requires training data, quality tests, and hosting.                                    |
| Tracking token spend              | What you can see                                                      | Usage is labeled by workflow, team, and model                                       | Only a mechanism for tracking; does not lower spend itself.                            |
| Trim prompts, context, and output | Tokens per request                                                    | Requests carry content the model does not need                                      | Cutting a needed instruction can lower quality                                         |
| Route tasks to smaller models     | Price per token                                                       | A smaller model passes your quality tests for the task                              | Needs a test set and a router, and the provider's price per token still applies        |
| Cache repeated prompt content     | Price of repeated input                                               | Requests share a long, stable beginning and arrive within the cache lifetime        | Cache writes can cost more than standard input, and output gets no discount            |
| Batch non-urgent work             | Price per token                                                       | Results can take up to 24 hours                                                     | No streaming, and some providers’ zero data retention policy does not cover batch jobs |
| Cap agent runs                    | Worst-case cost of one run                                            | You can set step and token limits per run                                           | You can set step and token limits per run                                              |

### What Is a Custom AI Model?

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 a defined set of tasks. [Fine-tuning](https://adaptionlabs.ai/blog/why-fine-tuning-fails), which trains an existing model on new data, is one common way to build one.

### Comparison: General-Purpose Model vs Custom Model Costs

|                      | General-purpose model through an API                    | Custom model                                    |
| -------------------- | ------------------------------------------------------- | ----------------------------------------------- |
| What you pay for     | Tokens at the provider's price                          | Training, then the compute to run the model     |
| Who sets the price   | The provider                                            | You and your hosting provider                   |
| Scope                | A broad range of tasks                                  | One defined set of tasks                        |
| Setup                | An API key                                              | Training data, training, and testing            |
| Where your data goes | To the provider with each request, under its data terms | Stays in your environment if you host the model |

### Why Does a Custom AI Model Cost Less to Run?

A custom model costs less to run because each request needs less compute, and the company stops paying a provider’s per-token price for work that does not need a large model. Three things drive the difference:

- Less compute per token. A model built for one set of tasks can have far fewer parameters than a general-purpose model, so each token takes less compute to process.
- A cost you set. Instead of a provider’s price list, the company pays for hosting compute, which follows model size, volume, and hardware. A provider’s repricing does not change it.
- Shorter requests. Behavior trained into the model can replace long instructions that otherwise travel with every request, which cuts input tokens.

Model choice moves cost a lot. [Ramp](https://ramp.com/blog/ai-token-cost-for-businesses) observed a 20-times difference in effective price per million tokens between the lowest-cost and a premium model in April 2026.

A smaller model can still do routine work well. In Adaption’s [September 2026 research](https://adaptionlabs.ai/blog/a-better-harness-can-unlock-smaller-models), A Better Harness Can Unlock Smaller Models, changing only the agent harness raised Qwen3.6-27B’s criterion pass rate on the Harvey Legal Agent Benchmark from 67.10 percent to 85.92 percent across 140 held-out tasks, with the model’s weights unchanged. Post-training then raised it to 88.03 percent, close to frontier performance on this benchmark. Most of the gain came from the harness, with post-training adding about two percentage points.

In the [July 2026 study](https://arxiv.org/abs/2607.08938), Better Harnesses, Smaller Models: Building 90% Cheaper Agents via Automated Harness Adaptation, researchers found related results across seven business tasks: in one case, a small model with an optimized harness recovered 89.7 percent of a large model’s performance at 4 percent of the cost.

On routine work, the choice is not between a large model you don’t own or a weak one you do. A smaller model can work reliably well at a fraction of the cost.

### Which Workloads Are Worth Moving to a Custom Model?

Move a workload to a custom model when it is routine, the volume is steady and high, training examples exist or can be generated, and you can test quality. Four questions to consider:

1. Is the work routine and well-defined? Bulk extraction, classification, summarizing to a set standard, policy checks, contract clause extraction, and resume parsing all fit.
2. Is the volume steady and high? Compare the workflow’s monthly token spend with the one-time cost to train a model plus the monthly cost to host and maintain it.
3. Do you have training examples? A model is only as good as its data. If you have internal data, such as logged requests paired with approved outputs or corrections to an agent’s past mistakes, the [Adaptive Data platform](https://adaptionlabs.ai/adaptive-data), Adaption’s real-time data optimization platform, optimizes, expands, and generates AI training data across 242 languages and localizations, so your data is ready to train on. If you have no data, [Invent a Dataset](https://adaptionlabs.ai/invent-a-dataset), Adaption’s synthetic dataset generation platform, builds a training-ready dataset from a description of the behavior a model should learn, with no example data needed.
4. Can you test quality? A test set of real requests with a pass mark set before training makes the comparison fair, and it shows whether errors come from the model or from the tools and context around it. Check the setup before concluding a model is too weak, and keep the larger model where the remaining errors reflect capability limits.

### How Easy Is It to Build Your Own Model?

[Building a custom model](https://adaptionlabs.ai/blog/best-platforms-to-train-ai-model-easily) takes four steps. [Automated data preparation](https://adaptionlabs.ai/blog/best-ai-training-data-platforms) and training cover much of the technical work, and the team chooses the task, supplies or approves examples, and tests the result against its own requests:

1. Define the task and a pass mark.
2. Prepare training examples.
3. Train the model against the outcome you set.
4. Test it on real requests and host it.

### Prepare Training Examples

The [Adaptive Data platform](https://adaptionlabs.ai/adaptive-data) is Adaption's real-time data optimization platform, which optimizes, expands, and generates AI training data [across 242 languages and localizations](https://adaptionlabs.ai/blog/expand-your-world). Adaption reports an average 82 percent increase in data quality across early deployments on average.

If you have limited or no data, because of privacy regulations, sensitive records, or a task with no history, [Invent a Dataset](https://adaptionlabs.ai/invent-a-dataset), Adaption’s dataset generation product in the Adaptive Data pillar, builds a training-ready dataset from a description of the behavior a model should learn, with no examples needed. Invent a Dataset covers specialized and regulated domains such as healthcare, as well as enterprise software, customer support, and low-resource languages.

### Automate the Research Loop With AutoScientist

[AutoScientist](https://adaptionlabs.ai/auto-scientist), self-improves and automates the full research loop behind model training and alignment. It adapts to what you are building, across a variety of industries including medical, technology, finance, and enterprise. Your team sets the outcome, and AutoScientist co-optimizes the model’s training data and training recipe until quality converges on that goal.

A smaller model replaces a provider only if it passes your quality tests, and reaching that quality has traditionally meant a team of specialists tuning each training run. Evaluations have shown that a model trained with AutoScientist’s settings beat an untrained model 64 percent of the time. This means a company can reach competitive quality without hiring that expertise.

AutoScientist also removes manual work: in a preliminary experiment with [AI Singapore](https://adaptionlabs.ai/blog/ai-singapore-sovereign-ai-pilot) on a small base model, AutoScientist saved 56 engineering hours of manual hyperparameter optimization. Those hours are part of what it costs to build a custom model, so saving them lowers the point at which one pays for itself.

Each step of the build has a product behind it. The Adaptive Data platform prepares the data a company already has, Invent a Dataset creates data where none exists, and AutoScientist runs the research loop that turns that data into a model the company owns. Because each step feeds the next, a team moves from a workload to a tested, owned model with fewer handoffs and less specialist effort. That shortens the time before a custom model can start replacing per-token spend.

### How Do You Reduce Token Spend if You Don’t Build A Custom Model?

Six options lower spend on the models a company already pays for. Each lowers the tokens used or the price per token and leaves the provider's pricing and terms in place.

### Track Spend by Workflow

Token spend tracking records the tokens each workflow, team, and model uses and what they cost. It lowers nothing by itself, but it shows which workflows drive the bill. Provider reports arrive after the fact, so a looping agent can run up cost before anyone sees it.

### Trimming Prompts, Context, and Output

Trimming removes content the model does not need and limits how much it writes back. Tokens concentrate in the system prompt, retrieved documents, chat history, and answer length. Send only what the task needs, request short structured output when software reads the result, and use ordinary code for fixed tasks such as reading a date from a field. Cutting a needed instruction can lower quality, and prompts grow back as teams add instructions.

### Routing Tasks to Smaller Models

Model routing sends each request to the lowest-cost model that does the job well: classification or formatting goes to a smaller 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. The router is another system to re-test when prices, models, or tasks change, and every request still pays the provider's price per token.

### 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. Cache reads cost a fraction of the standard input price, and the discount varies by provider.

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, 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 submits a group of requests together and returns results later at a lower price. Frontier model providers often charge 50 percent of standard prices for batch jobs, but results can take up to 24 hours, so batching suits work that can wait, such as nightly summaries or evaluation runs. It also has restrictions: some providers do not support streaming, and zero data retention may not apply to batch jobs, 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 sets a limit on the steps and tokens each run can use, plus a spending limit for each team that runs agents. Each step is billed and the agent chooses how many to take, so a cap stops a looping agent early. Caps control the worst case, not the average: an agent under the cap can still waste tokens on every run, a cap set too low cuts off legitimate long tasks, and a cap set too high does little. Choosing a cap requires knowing how many steps a normal run takes.

### What Is The Best Way To Reduce Your Company’s AI Token Spend?

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. Other strategies, like capping agent runs or batching non-urgent requests can help on workloads that stay on a general model, but your company is still at mercy to a provider’s pricing and terms. A custom model changes the terms themselves, and Adaption makes this path practical.

[Adaption for Enterprise](https://adaptionlabs.ai/enterprise)[Talk to an AI Expert](https://adaptionlabs.ai/book-a-call)

Author

[Adaption Team](https://www.linkedin.com/company/adaption-labs/)

Date

Oct 19, 2026