Rising AI costs: why you should pick the right model for each task
Large companies are switching to cheaper AI models. Here is what a small business can learn from this, and how to cut AI costs without losing quality.
What happened
According to Bloomberg, the gross margin of the legal AI company Harvey fell from roughly +50% to -50% by June, as token consumption grew 20-fold. In August Harvey launched its own model built on Moonshot's open model Kimi K3, and the margin turned positive again. Bloomberg also named Abridge and Ramp as other start-ups following the same path.
Financial Times reports that open models now handle 40% of AT&T's AI work. Most AI spending still goes to the closed models of the large labs, but big buyers now see the model as something that can be swapped out if the price is not right. The seller side is changing too: according to The Information, Anthropic is ending corporate discounts (typically around 15%) once a customer has used up the volume set out in the contract, after which the contract has to be renegotiated or the full price paid.
What this means for companies in Latvia
Most companies in Latvia do not use AI on the scale of Harvey or AT&T. The principle is the same, however, and for a small business it matters even more, because every unnecessary expense is felt immediately.
AI costs usually do not jump suddenly. They grow quietly: first one employee starts using AI for email replies, then document processing is added, then customer requests are automated. If all of this work goes through a single, most expensive available model, the bill rises with every new process. The Harvey case shows well that even a successful product can become unprofitable if consumption grows faster than revenue.
At the same time, the market is becoming cheaper and more diverse. On 7 October Anthropic released Claude Haiku 5.5, which, according to the company, costs about a tenth of the Haiku 4.5 price for most requests, around 75% less on average. It is intended for high-volume work such as customer support, summaries and data retrieval. On 6 October Mistral published a preview of Large 4, and once the open weights are released at the end of October, companies will be able to run it on their own servers and keep sensitive data in-house. For companies that care where customer data is stored, this is a significant option.
In our view, the lesson for a small business is the same as for the large ones: the smartest route is the model that fits the task, not always the most expensive one.
How to apply this in practice
Sort tasks by complexity
Most everyday AI work is routine. A cheap, fast model is enough for it:
- sorting emails - whether it is an order, a complaint, an invoice or a question;
- extracting data from documents - invoice number, amount, date, supplier;
- common customer replies - opening hours, delivery terms, order status;
- short summaries - the key points of long correspondence or meeting notes.
It is worth keeping the expensive model for complex tasks: contract analysis, non-standard customer situations, comparing multiple documents, or texts where every nuance matters.
Example: before and after
Before. In a company where enquiries arrive by email, every message is sent to the most powerful model. It reads the message, identifies the topic, extracts contact details and prepares a draft reply. The result is good, but every simple message costs as much as a complex one.
After. A cheap model first identifies the type of message and extracts the basic data. For standard questions it also prepares a reply from approved templates. Only if a message is unclear, contains a complaint or requires judgement is it handled by the more expensive model. Quality stays where it is needed, while most requests are processed at lower cost.
Build so that the model can be swapped
The AT&T example shows that large buyers treat the model as a replaceable component. A small business should also build its solution so that changing the model does not mean rebuilding the whole system. In practice this means:
- the connection to the AI model is a separate layer, not scattered throughout the system;
- instructions (prompts) and templates are kept in one place and are not tied to the quirks of a single provider;
- there is a small test set of real examples, so a new model can be checked before switching.
This approach lets you use new, cheaper models as soon as they appear, and avoids dependence on one vendor's pricing policy. The Information's report on discounts is a reminder that terms can change even during a contract.
Track the cost per task
The subscription price says little on its own. It is far more useful to know what one processed email, one scanned invoice or one customer conversation costs. Once these figures are visible, it becomes clear which processes are expensive, where a cheap model would be enough and where the most powerful one is really needed.
Where to start
You do not have to rebuild everything at once. A few steps are enough:
- List all processes where you already use AI, and note which model is used in each.
- Assess complexity. Which tasks are repetitive and simple, and which require judgement and careful evaluation?
- Choose one routine process and try it with a cheaper model, comparing the results with the current ones.
- Start recording the cost per task, so that decisions are based on data rather than impressions.
- Check the architecture. If changing the model would require major rework, that is the first thing to sort out.
If you want to understand where AI costs in your company can be reduced without losing quality, contact AI-Lab Space. Together we will review your processes, recommend a suitable model for each task and help build a solution that can be adapted as the market changes.