Russian AI services in 2026: three months of operation in figures

Between May and August 2026 we made 30,449 real calls to YandexGPT in production — on live user requests, not in testing. This is not a market survey but a report on three months of operating one service, and no such data exists publicly.

The response time distribution:

17.2% of calls answer faster than two seconds and 0.7% take longer than ten. The tail is short: median to p99 spans only two and a half times, where for ordinary web requests we measure a tenfold spread.

Empty responses number 43 of 30,525, or 0.14%. For three months of continuous operation that is a good figure.

Mean latency rose from 2.2 to 4.0 seconds across three months. It is easy to read that as "the service got slower", but the data says otherwise:

Latency rose 82%, answer length 77%. Per token the speed holds at about 14 ms and has not moved in three months, while load grew eightyfold.

The practical conclusion: a language model’s response time is predicted by answer length, not by service load. Want it faster, ask for less. A cap on output tokens works more reliably than any timeout.

Our Russian requests go to a Russian provider and English ones to a foreign provider. This is neither a price optimisation nor a quality judgement but a data-handling requirement: transferring user information abroad is regulated, and the split removes the question at the root.

There is a technical side too, which we checked. From our Moscow server:

The network path to all three is short — tens of milliseconds. But the foreign endpoints refuse a request from a Russian address. So the English side of our system works through an intermediate node outside the country, and part of its response time is network path rather than model work. The median there is 5,489 ms against 3,793 for the Russian provider, and the two must not be compared directly as "model against model".

No market shares for Russian AI services exist. No regulator, no industry body and no independent researcher publishes how many companies use which service. Everything appearing in roundups is either a retelling of marketing claims or invention.

We did not test other providers’ endpoints at guessed hostnames. A negative result on a wrongly guessed host means only that the name was wrong, not that the service is down. Publishing such a thing as "the service does not work" is a common and crude error.

Our data describes one service and one kind of load. We ask the model to summarise a technical check — short structured input, short output. A chat with long context, or code generation, would show an entirely different profile.

What this data does give: a benchmark for embedded-model response time on Russian infrastructure. If you are designing an interface, assume four seconds for a typical answer and ten for a bad case — and show the user that, rather than spinning silently.

To see how it looks on our side, try any check page — the DNS lookup or the SSL checker, for instance.