AI Pulse · Under the Radar

A developer has published a firsthand account of moving a 35kb production prompt away from Anthropic Opus and GPT APIs to a self-hosted setup using Ollama. The useful part is not a claim that one route is always cheaper or better. It is the practical friction that appears when a system already works in production and you try to replace the model layer beneath it. You can read the full account in the original writeup.

For an owner, this is a reminder that AI cost decisions are rarely just API price comparisons. A large prompt can be tied to model behaviour, context handling, reliability expectations, hosting capacity and the work needed to test whether the replacement produces acceptable results. The upfront bill for a self-hosted model may look attractive, but the real decision includes engineering time, operational ownership and the cost of errors if outputs change.

The post has drawn 89 points and 40 comments, which suggests this is a live concern for teams thinking about their dependency on major AI providers. If your company has important workflows built on one provider’s models, ask a simple question now: do we know what would break, what it would cost, and how long it would take to move?

This is the kind of thing we build into an AI command centre, tracking model dependencies alongside the business processes that rely on them.