The Price of Perfection

After ElevenLabs, Asya's brain became OpenAI Realtime — and for several weeks it felt like the search was over. Conversations flowed, scenarios closed, there was almost nothing to fault. There was something else to fault: the number on the bill. This is a dispatch about being defeated not by technology but by arithmetic.

Let's start with the honest part: OpenAI Realtime is a strong engine. On our test runs it handled the scenarios more reliably than anyone we worked with before or after. The prompt moved out of a cloud dashboard into a repository; edits to the character became code with history and rollback — exactly what we'd left the previous engine for. Life got noticeably faster.

It had its quirks, of course. The cheaper, smaller model stubbornly slid into Ukrainian during Russian conversations — no matter how many rules we wrote; reliable bilingualism came only from the larger model, twice as expensive per token. A fork in the road: either the agent confuses languages, or you pay double for it not to.

There was one story we still tell as an anecdote. In the web widget, where Asya answers right on the site, the first half-second of her own greeting bounced from the speakers back into the microphone. Asya heard herself, mistook herself for the caller, politely fell silent, started over — "Good day!.. Good day!.. Good day!.." — round and round, a duet performed by one person. We cured it with a pause: for the first seconds of the greeting she is now deaf to interruptions.

And once, she simply invented the price list. Instead of checking with our own systems, she confidently quoted a caller speeds and prices the provider had never offered. The tool sat right there, available — she didn't use it. That was the first time we seriously considered that "sometimes doesn't call its tools" is not a minor line in a bug list but a separate axis of quality, one that engines must be measured on as strictly as on sound. Remember that thought — it will come back later in our story, in the dispatch about Grok.

But all of that was fixable or tolerable. What proved unfixable was the arithmetic. When a month of live-call statistics had accumulated, we calculated the effective cost of a minute of conversation — with every cache, discount and optimization we could squeeze. It came to twenty-two point six cents per minute. The alternatives we were running in parallel by then delivered comparable quality at five cents. Five times cheaper.

Five cents versus twenty-two is not "saving a little." It is the difference between "the agent pays for itself" and "the agent is an expensive toy." For one provider with its call volume — noticeable; for a blueprint we intend to replicate across other businesses — a verdict. A minute of conversation is the unit cost of our product, and it must be predictable and low, or nothing else matters.

So we parted with this engine too — respectfully, without hard feelings, and not even completely: it is still on duty on one of the lines while the final migration is under way, and it remains our fallback for a rainy day. The quality never went anywhere. It's just that quality is half the equation.

And here is where our story turns into a proper circus. The next candidate came from the company with more data and more compute than everyone else combined — and merely to keep it alive on the line, we had to post five watchdogs around it at once. Gemini — in the next dispatch.