Two months with WorkBuddy: rerun count, not token price, is what really drives cost

Yesterday I scrolled past CAICT's "Large Model Token Service Performance Monitoring Report," the domestic normalized average price went from 6.8 yuan in the second half of 2025 to 13.46 yuan this year, nearly double. In the same news there's another number, in August the domestic model average TPS reached 78.23, double the 2025 average of 38.25. First-token latency is also rising, from 0.78 seconds to 1.26 seconds. I read those lines twice, but what came to mind was that afternoon last month organizing reimbursement forms.

I wrote about that, twenty-some invoices, PDFs plus phone screenshots, four amounts didn't match, 86 on the invoice exported as 68, and two had blank dates. At the time I thought the photos were crooked, retook them and re-uploaded, same old thing. Later I pulled those four out and processed them separately, one at a time, and then it was right.

That bug is another matter. What this piece is about is, at a moment when the average token price has doubled, what's the real difference between a finished tool like WorkBuddy and me writing my own script to call the API.

I did a comparison myself once. Same task, a club's activity records for one semester, thirty-some documents plus two tables, to be summarized into one table with date, topic, participant count, budget. Route A is me using Python to call the API, writing the prompt myself, parsing the returned JSON myself. Route B is WorkBuddy, just throw the files in.

On recognition quality, the two really don't separate. WorkBuddy is also a model underneath, when the document format is regular, fields get extracted pretty accurately on both. The difference comes after. When Route A errs, I can see the raw return, know whether the model missed something or my regex was wrong, change one line of code and rerun. When Route B errs, all I see is the result table, I have to guess. Those four wrong-amount invoices, I guessed three times before I got the cause right.

But Route A's cost is completely uncontrollable. One document I have to split into several chunks to feed, with retries in between, tokens burn per call. With WorkBuddy I have no idea how many times it called the model internally, I only know it finally handed me a table. Sounds like a downside, but from another angle, precisely because I can't see it, I don't rerun anxiously over and over.

In my testing, a three-thousand-word activity record takes maybe tens of seconds per run. If I rerun five times because the result is wrong and I changed the prompt, the cost is five times. That 13.46 yuan from CAICT is a normalized average, for a single task it might be a few cents, but multiplied by five or ten times, the math isn't the same. The real impact of token unit price doubling on ordinary office users is far smaller than the waste from repeated reruns.

So the way I use WorkBuddy has changed too. Before I always wanted to state the whole requirement at once, categorize by date, topic, participant count, budget, keep budget to two decimals, unify dates to 2026-XX-XX format, truncate topics over ten characters. Result: it often loses one thing chasing another, budget format right, dates messed up. Later I learned my lesson, one thing at a time. First only do field extraction, export and I adjust format myself in Excel. Slower, but reruns dropped from an average of three or four to just over one.

Another pitfall is field names. I wrote a whole piece on this last month, and I have to say it again, because it's the easiest place to trip on WorkBuddy. My activity record table, the header says "participant count," I wrote "count" in the requirement, the whole column exported empty. It aligns strictly by field name, one character off and it treats the field as nonexistent. I think this design is right, strict beats guessing, but it's especially unfriendly to beginners, because the error is silent, all you see is a blank column.

My friend tripped on this last week too. He was organizing club reimbursements, the header said reimbursement amount, the requirement only said amount, whole column empty. He fiddled for over an hour, thought the file was too big to process.

WorkBuddy's biggest problem is the black box. It doesn't tell you what step it's on, doesn't tell you which files failed, doesn't tell you how many times it called the model internally. Throw a dozen-plus files in at once, one gets stuck in the middle, you can only wait, and when it's done you find the result table is missing two rows, and you have to go back and figure out which two. This experience doesn't matter with few files, but past twenty it's really annoying.

By comparison, my own script is clumsy but every step has logs, failures can be traced to a specific file. So my division of labor now is, jobs under ten files go straight to WorkBuddy, over twenty, or requiring reproducibility and auditability, I go back to writing scripts. Financial reports and things that need a paper trail, I won't hand to a tool that won't tell me what happened in between.

There's also a hidden cost the report didn't mention. Context length is expanding, among monitored models the share with 1M context has reached 47.8%, 200K and above totals 95.7%. Tools can now stuff a whole long document in without chunking. Sounds like a good thing, but for a user like me, the more it stuffs, the harder it is for me to judge which part it actually read. Once I threw a whole semester of meeting minutes in, asked it to summarize resolutions, and it counted a resolution from last semester too. The file was misplaced by me, but it gave no warning at all.

WorkBuddy is suited to batch-cleaning dirty work with clear structure, not suited to accounts that need a paper trail, precise numbers, or after-the-fact accountability. For the former it's much faster than hand-written scripts, for the latter it's much more dangerous than hand-written scripts.

Now price. That CAICT report says foreign typical models average 117.8 yuan this year, domestic 13.46 yuan, nearly nine times apart. This gap won't disappear short term. For a product like WorkBuddy, running on domestic models, its cost structure is advantaged. But I don't think this advantage passes to users, because office tool pricing logic never looks at token cost, it looks at how much you're willing to pay for the time saved. So don't expect WorkBuddy to get cheaper just because models get cheaper.

Back to that reimbursement afternoon. My approach now is, first throw all invoices in for a round of recognition, export the table, then check amounts and dates row by row myself in Excel. Checking takes about twenty minutes. Those twenty minutes can't be skipped, because I don't know which one will be wrong. WorkBuddy compresses a two-hour job into forty minutes, the remaining twenty minutes is my insurance.

I saved that CAICT report, next time prices change I'll pull it out and compare. I'll keep using WorkBuddy the way I do now: one thing at a time, check it myself after export.