Model choice is a scoping decision, not a religion. OpenAI, Anthropic and the open-source ecosystem each earn their place in different workflows, and a single system often mixes them. Here is the frame we use when deciding.
Hosted model APIs
Hosted APIs — like OpenAI and Anthropic — are the default for language-heavy features. They bring frontier capability without you operating infrastructure, they scale with usage, and they improve as the provider ships new versions. The trade-offs are data leaving your environment, per-call cost that grows with volume, and dependence on a provider's API and roadmap.
Open-source models
Self-hosted open-source models earn their place on three tests: data that must not leave your environment, call volumes where per-token pricing changes the economics, and the need to control the stack end to end. The cost is yours to carry: the infrastructure, the evaluations, the updates and the operator skill to keep the model current and behaving.
The decision hinges on the workflow
Ask what the model is doing. Extracting fields from an invoice, summarising support threads and answering from internal knowledge each impose different demands on data handling, latency, cost and accuracy. A document-processing pipeline that runs inside your network might use a self-hosted model for extraction while a customer-facing assistant uses a hosted API for reasoning.
Two practical notes. First, capability moves fast in both directions — a model that was the obvious choice six months ago may not be now, so write the integration so the model is a swappable component, not a hard-wired dependency. Second, whatever you choose, the workflow needs an evaluation set: a batch of real inputs with known-good outputs, so you can measure whether a model change actually improves the system. That is the discipline that keeps model choice honest.
Model-agnostic means we make this call per workflow and name it in the quote. If you want a second opinion on your stack, start the conversation.