Anyone who builds AI functions into their own products or has development teams working with model support has a new line item in their cost accounting: billing per request with the model provider.This line item is currently in flux.According to reports, OpenAI is specifically poaching customers from Anthropic and granting discounts of up to 90 percent for this.Companies such as Rippling, Replit and DataCamp report corresponding cost reductions.
The background is a battle for market share.Anthropic's growth has reportedly slowed, OpenAI is buying market share through price – and paying for it with its own margin.At the same time, Chinese providers such as Deepseek are pushing the price level down from below.The result is a market in which the costs for comparable performance fall faster than most budget plans can reflect.
Why this is relevant for projects with an AI component
In classic web projects, ongoing costs are easy to calculate: hosting, maintenance, licences.As soon as a product triggers model calls – for example for text generation, classification or search – a cost category arises that scales with usage.As the number of users rises, the bill rises with it.
If prices fall, this bill becomes cheaper.At first that sounds merely pleasing.In practice, however, it shifts the feasibility threshold: use cases that were abandoned a year ago for cost reasons may be viable today.Anyone who has made such decisions should review the assumptions instead of carrying them forward.
Discounts are not a permanent state
At the same time, caution applies.Price reductions of this magnitude arise in a competitive situation and not from lower operating costs.If a provider buys market share at the expense of its own margin, that is not a stable permanent state.Either production costs fall accordingly, or prices rise again once the market positions are settled.
For the calculation, this means: a special-conditions offer is a good basis for a pilot project, but a bad basis for a business model that is meant to hold up over years.Anyone who builds their contribution margins on a discount has a risk on the balance sheet that lies outside their own sphere of influence.
Provider independence becomes an architectural task
The actual consequence lies in the technology.If prices and performance ratios shift several times a year, switching model providers must be a manageable process – and not a new development.
That is not achieved on the side, but through deliberate design decisions:
- Abstraction layer: The application does not talk directly to a provider, but to its own internal interface.Behind it, the respective model sits in a replaceable way.
- Prompts as configuration: The instructions to the model do not belong firmly in the program code, but in a central, versioned place.Only in this way can they be adapted when switching providers without rebuilding the application.
- Measurable quality: A set of defined test cases with expected results shows whether a cheaper model is sufficient for the specific use case.Without such tests, the comparison remains a gut feeling.
- Cost transparency per function: Which part of the program causes which share of the model costs?This allocation is the prerequisite for optimising in a targeted way instead of saving across the board.
- Data protection and location: Switching providers is also a legal process.Where data is processed and under what conditions must be checked before the switch, not after.
Not every task needs the most expensive model
Another lever lies in-house, independently of the price war.In many applications, all requests run through the same, most powerful model – including those that require a simple assignment or a short summary.A tiering by task type often lowers costs significantly without users noticing a difference.
In addition, there are classic measures: caching recurring requests, shortening input texts to what is necessary, not regenerating results multiple times.Such optimisations work regardless of which provider currently offers the best price – and they remain effective when the discount phase ends.
Review contracts and dependencies
Anyone who accepts a discount offer should know the consideration.In practice, minimum purchase commitments, terms or exclusive ties are common.The longer the commitment, the lower the benefit of falling market prices during the term.A shorter contract term at a somewhat worse price can be the better choice in such a volatile market.
It is equally worth looking at operational stability: if a provider buys market share through price, it is not guaranteed that capacity, response times and support remain at the usual level.A second provider kept ready in the background is not only price insurance but also failure insurance.
Classification for web and software projects
For ongoing projects, the price war above all means: model selection is no longer a one-off decision, but a decision that should be reviewed regularly.Technically, this only pays off if the connection is built to be replaceable from the start – retrofitting this is expensive.Anyone planning an application with an AI component today should include the abstraction layer and a simple test suite for model quality as a fixed part of the project scope, not as a later extra.