OpenAI, Anthropic or Google: choosing a model provider without betting the business

OpenAI, Anthropic and Google all ship excellent models, leapfrog each other quarterly, and price aggressively. Treating the choice as a marriage is the mistake; treating it as plumbing is the win.

The current shape of the field

OpenAI has the broadest ecosystem and the most developer familiarity — the default choice, in both the good and lazy senses. Anthropic's Claude models have a strong reputation for long-document work, careful instruction-following and natural prose, and the company has pushed hard on agentic tooling (we compared them in detail here). Google's Gemini models are priced sharply, integrate deeply with Workspace, and their cheap tiers are excellent for high-volume work. All three offer UK/EU-compatible commercial terms with no training on your API data.

The architecture that keeps you free

• One thin routing layer. Every AI call in your systems goes through a single internal interface. Swapping provider becomes a configuration change, not a rebuild.
• Task-level routing. Classification and extraction go to cheap fast models; drafting to a mid-tier; the genuinely hard reasoning steps to a premium model. Most systems we build spend 80% of their volume on models costing pennies.
• Your own eval set. Twenty to fifty real examples from your business — your emails, your documents, your edge cases — with known good answers. When a new model ships, run the set, read the failures, decide in an afternoon. Without this you're choosing on vibes and blog posts.
• Prompts as assets. Keep prompts versioned and provider-neutral where possible; they're business logic, not throwaway strings.

What lock-in actually looks like

Rarely the API — those are easy to swap. Real lock-in hides in provider-specific features with no equivalent elsewhere, fine-tuned models you can't port, and data living inside one vendor's tooling. None are fatal, but each should be a deliberate trade, not an accident.

The uncomfortable truth: for most business automation, the difference between the top providers matters far less than the quality of the system around the model — the data it can see, the approval points, the measurement. Which is convenient, because that part you get to own.

← All notes