Anthropic and OpenAI change models and prices under you constantly. How exposed is my board process to your AI vendors' whims?

The full executive answer

Less exposed than any single-model product, by construction — the exposure is managed at three layers. Layer one, no single point of dependence: the AI gateway abstracts the provider, Claude is primary, GPT-4o is the standing fallback, and both the primary and fallback model identities are configuration, not code — a model deprecation is an environment-variable change and a regression run, not a rewrite. Layer two, and this is the part most AI products skip: model-independence of behaviour. Every AI output must conform to a typed, machine-validated schema, and the evaluation harness in CI runs golden fixtures through every one of those contracts on every push. When a vendor swaps a model version, we do not discover breakage from customers — the contracts either hold or the gate fails, and the corrective-retry mechanism absorbs minor behavioural drift at runtime.

Layer three, economics and continuity: the response cache means already-generated insights survive any provider outage or price spike at zero marginal cost, and per-response records of provider, token counts and fallback usage give us the actual cost telemetry to renegotiate or re-route deliberately rather than react blindly.

The residual honest exposure: a simultaneous outage of both frontier vendors would pause new generation — cached and exported work survives — and a coordinated price rise across the industry is a margin question, not an architecture question. Managing exactly this class of third-party AI dependency is what the NIST AI Risk Management Framework asks of AI-dependent operators, and the two-vendor, contract-tested, cached architecture is our implementation of it.