What's running today, and what isn't yet

Every platform's website describes the finished version. Here's the actual state, because you'll find out anyway and it's better you hear it from us.

What happens when a model provider goes down

Worth knowing because it fails in a misleading way. Spoken calls do not depend on our language model at all — the voice service supplies its own — so a call carries on sounding perfectly normal while chat, post-call reports and answer quality stop. The healthy-looking half is the loud one.

Against that there is automatic failover. Several providers are configured in order, a provider that returns a quota or billing error is recognised as permanently failing and skipped immediately rather than retried, and one that keeps failing is stood down for a cooling-off period instead of being asked again on every request. Traffic moves to the next provider without anyone intervening.

The honest limit is that failover needs somewhere to fail over to. Two configured providers that have both run out of credit is not redundancy, it is one outage twice — and that is a billing question rather than an engineering one. Production runs with more than one funded provider for exactly this reason.

Predictive dialling is deliberately switched off rather than unfinished: it needs the two-second information message and answering-machine detection before it can be used lawfully, and we would rather ship it late than ship silent calls.

And the thing you should judge us on. None of the above tells you how it sounds on a bad connection, with an accent it has not heard, in a room with noise in it. That is where voice AI actually succeeds or fails, and it is not something a website can settle. Call the demo line and interrupt it, talk over it, use a phrase it will not expect. If it holds up, the rest of this page is worth reading twice. If it does not, nothing else here matters.