Industry Impact2 min read

AI Office Work Now Costs Cents, So Checking It Is the Skill

By , Senior AI ConsultantPublished

Tencent's new Hy4 flash model does everyday office jobs for about a twentieth of the cost of its flagship, which moves the valuable work from doing the job to choosing the model and checking the result.

Tencent's new AI model, Hy4 flash, is built for the plain middle of office work: slide decks, document edits, spreadsheet repairs, audio and video clean-up and reading images. On Tencent's own set of typical office tasks, a finished job costs about a twentieth as much as on its flagship Hy4. That makes it a good case for a question every team that pays for AI will face: which jobs need the best model, and which only need a cheap one that someone checks?

The cheap model is weaker, and that is fine

Hy4 flash scores 42.1 on that task set against 51.4 for Hy4. That is about 82% of the score for about 5% of the cost per job. Tencent itself keeps long software-engineering projects on Hy4 and points everything else to the cheap model: research, documents, slides and simple screen work.

The saving is bigger than the price cut alone. Per-token prices are about 90% below Hy4's, but the cost per job falls 95%, because the model was trained to use about half as many tokens, the small chunks of text it writes, to finish the same job. A bill is the price of a token times the number of tokens, and both went down. Alibaba's model showed the same pattern: it got 67% cheaper per task on a 25% price cut.

When Tencent's own staff rated results without knowing which model made them, satisfaction was almost level with Claude Haiku 5.5, and a job cost RMB 0.26, about 4 cents, with a median finish time of 6.7 minutes. The model is free in Tencent's WorkBuddy assistant until November 30, and Tencent plans to release its weights.

Run the cheap model first, and pay for the strong one only on failures

For example, the office manager of a small distributor hands AI 500 routine jobs a month: the weekly sales deck, the supplier price sheets, the contract summaries. At 4 cents a job that is about $20 a month. At twenty times the price it would be $400, the difference between something nobody has to approve and a line in the budget.

The better setup uses both models. Every job runs on the cheap one first, someone checks it, and only the failures go to the strong one. If one job in ten fails, the average job costs about three times the cheap price instead of twenty, and the hard jobs still get the stronger model.

The check is the part that needs a person, and it is quick: totals that must match the source sheet, a deck whose numbers can be traced, a summary read against the one clause in the contract that matters.

Who gains when the arithmetic is free

After the spreadsheet arrived, the clerks whose work was the arithmetic lost their jobs, while accounting jobs grew, because the people who understood the business could suddenly test twenty plans where they once tested one. Expect the same split in AI office work. Someone who only reformats decks is exposed. The person who owns the weekly report and knows when a number looks wrong can now ask for twenty versions of an analysis for under $1.

Whoever in an office can look at a finished spreadsheet and know within a minute that it is right is the person a model like this makes more valuable.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable