Why we went model-agnostic. The bet that almost killed us

Early on we had a real fork. Build deep on one model (GPT-4o was the obvious pick) or build a model-agnostic routing layer.

Going model-agnostic was painful:

- Slower initial development

- Three times the testing surface area

- Customers asked 'why not just use GPT?' constantl

Then two things happened in 6 months:

- OpenAI pricing for some workloads got beaten by Claude on quality and by Llama on cost

- Three of our biggest deals were won because we could deploy local Llama on-prem when others couldn't

The lesson, picking the LLM at the platform level is a 12-month bet. Picking it at the agent level is a Tuesday decision.

We're AI Hive, this is the bet that made us what we are.

What strategic call did you make early that you'd defend even harder today?

Would love to hear your honest opinion on this.