Cheaper isn't a reason to switch AI models

OpenAI is buying customers from Anthropic. The Information reported that it has been offering steep discounts to some of Anthropic’s biggest accounts, and The Decoder summarised the numbers. Replit moved after OpenAI took more than 30% off a contract worth over $100 million a year. DataCamp’s chief executive said the new pricing cut his costs by 80 to 90%. At Rippling, chief executive Parker Conrad said OpenAI’s share of the company’s coding work went from 16% to 52%, while Anthropic’s fell from 57% to 18%. One US bank had its ChatGPT seats doubled to 500 at no extra charge, and now pays under $7 a user against the usual $20 or more.

Some of it is genuine. OpenAI’s newer models are cheaper to run, and the prices have come down with them. Some of it is OpenAI giving up margin to win market share, which is what a well-funded challenger does.

Would we move?

No, not on price alone.

We use both, for different jobs. OpenAI is very good at image generation and anything visual, and when the work is creative it’s often what we reach for. Anthropic is where we do the reasoning: the rationale behind a strategy, pitches, proposals, and building agents that do real jobs inside the business. On that kind of work it’s still the stronger of the two for us, and the difference shows in what comes out.

Discounts this size are aimed at companies buying one thing in bulk, mostly code. An agency isn’t buying in bulk. It’s buying the quality of an argument that goes in front of a client, and a cheaper model that writes a weaker rationale costs more than it saves the first time a pitch is lost.

What does switching actually cost?

The price is the visible part. The rest is everything you’ve built on top. Agents are tuned to how one model behaves, and the instructions and checks around them are written for its quirks. Move, and all of that gets redone and tested again, by the same people who were meant to be getting time back.

There’s also a reason to be wary of a discount a supplier is paying for out of its own margin. Prices set to win share tend to rise once the share is won.

How should an agency choose?

Pick the model by the job, not by the invoice. Run the work you actually do, a real proposal or a real agent task, through both, and judge what comes back. Then use price to decide between options that are close, not as the reason to leave the one that’s better.

Cost control matters, as I wrote when Accenture capped its staff’s Claude spending. But the cheapest model is only a saving if the work is as good. If it isn’t, you’ve moved the cost from the AI bill to the client relationship.