How Hunch works
A Hunch formula sends the text in your cells and your question to Jev, TypeSafe's System One model, and writes the answer back as a number or one of your labels. Jev returns judgments with probabilities instead of generating text, so every cell holds something you can sort, filter and count.
What happens when a formula runs
- The formula collects the texts in its range, drops blanks and duplicates, and checks its six-hour cache.
- The rest go to the Hunch server in requests of up to 40 rows, several requests in parallel.
- The server checks your key and credits, then asks Jev one structured question per row: yes/no for HUNCH, a choice for HUNCHPICK, a score for HUNCHSCORE, several yes/no questions for HUNCHMULTI.
- Jev returns a probability, or a probability for every option or level. The server charges one credit per answered row and sends the answers back.
- The formula writes the values into the cells and caches them, so recalculating the sheet costs nothing for six hours.
A 40-row request took about 1.7 seconds in our tests. In Google Sheets a single formula handles up to 1,500 unique texts within Google's 30-second limit on custom functions; the Hunch menu's Ask about selection handles bigger columns by writing values in chunks for up to six minutes.
Why a number instead of text
Most AI functions for spreadsheets send your text to a chat model and paste its reply into the cell. The reply is text: it can vary in wording, it needs cleanup before you can count it, and it's billed by the token. Jev is a System One model. It answers a defined question with a calibrated judgment, such as a probability of yes, one option from your list, or a position on your scale.
A judgment has three practical consequences in a spreadsheet. The column sorts. The labels are always exactly yours, so COUNTIF and pivot tables work on them directly. And the price can be fixed per row, because an answer does not vary in length.
Calibrated means the probabilities are meant to match how often the answer turns out yes: of many rows at 0.8, about 8 in 10 should be yes. Check it on a sample of your own rows before relying on a threshold.
What it is good at, and what it is not
- Good at judging meaning in short and medium texts: intent, tone, topic, fit, urgency, whether a condition holds.
- Not built for arithmetic, counting items in a text, or comparing dates. Keep those in normal formulas.
- Not a text generator. It will not write summaries, translations or emails. A chat model is the better tool for those.
- English works best.
What is sent and what is stored
The text of the cells a formula points at, and your question, go to the Hunch server and on to TypeSafe to compute the answer. Hunch does not store your text, questions or answers. It stores your key, its credit balance and the email used to buy it. The full policy is on the privacy page.
Questions
Which model does Hunch use?
Does Hunch train on my data?
Can I call it without a spreadsheet?
Run it on your own column
Copy the template (or use the Excel add-in), paste your key, point the formula at your data. Every formula here has its Excel spelling underneath. 100 rows free to start, then $29 for 5,000 rows. Credits never expire and there is no subscription.