Tencent's new Hy4 model can look at images, use a computer and work through long tasks, and it is on sale at about $0.90 per million tokens read and $2.70 per million written, the same prices as its August preview. A token is roughly three-quarters of a word. The price tag is the part of this news that changes what an ordinary office can do.
An assistant that watches a day reads it dozens of times
A model does not remember anything between requests. An assistant that tracks a project lead's messages, email, meeting notes and calendar has to be handed that whole record each time it looks, and to catch a slipping deadline before it matters it has to look often.
Take a record of 100,000 tokens, about the size of a short novel, checked 50 times in a working day, roughly every ten minutes. That is 5 million tokens read. At $0.90 per million, it costs $4.50 a day, close to $100 a month for one person. That is too much to leave switched on for a whole team.
Hy4's price for text it has already read is $0.045 per million, a twentieth of the normal price. The saving only works when each request starts with exactly the same text as the last one, and a work record does that, because new messages are added at the end and the rest stays the same. The day then costs about 23 cents to read. Add a few cents for the short notes it writes, and the whole thing comes to roughly $6 or $7 a month per person, or about $260 a month for a 40-person office.
The weak spots are not what this work needs
Tencent itself describes a gap with the top proprietary models on research-level maths, long coding jobs and hard questions about images. Almost none of that is on a project lead's list. Chasing deadlines, noticing that a meeting needs preparation and pulling the right message out of three weeks of chat are reading tasks, and a model that is one step behind the best handles them.
Accuracy still matters. On Tencent's own test the model made things up 7.2% of the time, down from 12.95% for the August preview. The sensible conclusion is to give a cheap model the work whose answers can be checked in seconds. A flagged deadline can be checked against a calendar. A fund report cannot be checked that fast, so it stays with a stronger model and a person who reads it.
For a business, that gives a simple way to split the work. Send the watching, sorting and reminding to the cheap model, and pay for the expensive one only on the few tasks where a wrong answer costs real money.
Where the model runs decides what it may see
An assistant that reads every message and document a person has sees everything, so where it runs matters as much as its price. Hy4 has about 770 billion parameters, which means a rack of servers to run it. Tencent has promised to publish its files in the coming days, and a large company with its own servers could then keep the record inside the building. Everyone else will reach it through a host such as Tencent Cloud or OpenRouter, and the record travels with every request.
So anyone picking one of these assistants for a team should ask where the record is sent before asking how clever the model is.