You run one model successfully
Our own benchmark found a single good mid-tier model is close to Pareto-optimal on general traffic. If that describes you and you have no governance requirement, a thin wrapper is the right call.
None of the pieces are hard. The difficulty is that the parts which look easiest are the ones that fail silently — a budget cap that is breachable under concurrency looks identical to one that works until the month it does not.
| Component | Realistic effort | Where it goes wrong |
|---|---|---|
| Provider abstraction | A week | Genuinely easy. Three adapters behind one interface. |
| Retry and fallback | A week, then months | The loop is easy. Knowing which failures are retryable, and revalidating policy before each attempt, is where it goes wrong. |
| Cost estimation | Two weeks, then forever | Prices change. Thinking models bill reasoning as output. An unpriced model must not read as free. |
| Policy composition | A month | Deny-overrides with monotonic tightening across five scopes. Easy to write, hard to keep correct as scopes multiply. |
| Budget enforcement | Two weeks | Trivial until concurrency. A counter-based cap is breachable by two simultaneous requests and nobody notices until the invoice. |
| Quality measurement | Months | A benchmark everything passes measures nothing. Building one that discriminates, and excluding mock runs and infrastructure errors from it, is most of the work. |
| Manifests and provenance | A month | Recording what ran is easy. Recording what was refused, and why, means threading rejections through the planner into storage. |
Our own benchmark found a single good mid-tier model is close to Pareto-optimal on general traffic. If that describes you and you have no governance requirement, a thin wrapper is the right call.
If model selection is a differentiator rather than plumbing, own it. Buying a control plane to then override its decisions everywhere is the worst of both.
Manifests, immutable versioning and residency constraints cost real engineering. If nobody will ever ask you what ran under which rules, that spend is not yet earned.
A build-vs-buy page written by a vendor is an advertisement. Read this one as an argument, not an assessment. The specific thing we think is genuinely hard to replicate is not routing — it is the discipline of never letting an estimate present itself as a measurement, and that is a property of how a system is built rather than a feature you can copy in a sprint.
Free tier, your own key, and a base-URL change.
Get started