Not a keyword filter — a model reads every email and keeps the three that matter

Here's the idea in one line: you've got a firehose — 200 emails by morning, a channel feed, 500 job posts, hundreds of messages. A plain keyword filter either misses what matters or drowns you in junk. Instead, a small smart model reads every item, scores it against your criteria, and shows only the top three — each with one line on why it matters.

And this isn't the bot that squeezes one article down to five lines, and it isn't the agent that watches the web waiting for one piece of news. This is your own stream, sorted for exactly you.

Why this just became possible

Running a good model over every one of two hundred emails used to be expensive — so people filtered by keywords. And keywords are a dumb filter: something "important" without the word "urgent" slips through, while an ad with the word "payment" gets dragged in. It all broke.

On July 9, 2026 OpenAI shipped Luna from the GPT-5.6 family — a "high-volume workhorse" at $1 per million input tokens, built for exactly the "classification, tagging, short answers that run millions of times" kind of job. Reading every email with a model now costs pennies. And you can filter by meaning instead of words: not "does it contain urgent" but "does this actually matter to me."

What you'll learn

• Scoring by your own criteria. Not "contains a word" but "rate 0 to 10 how important this is to someone waiting on a project reply and invoices due." The model judges by meaning.
• Structured output. You ask for a strict {score, reason} per item — then just sort by the number and take the top three. The answer becomes data, not prose.
• The cheap model on volume. You'll see on your own bill how $1 per million changes what you can afford: "read everything" used to be a luxury, now it's the default. And you'll learn how not to overpay.

A ready starter prompt

Don't ask for "filter my emails" — you'll get one lumpy summary. Give the criteria and a strict answer format per item.

The strong prompt leaves nothing to guess: you can see the criteria built for you, the strict {score, reason} per email, and the "don't invent" rule. What comes out isn't a summary but a sorted list you can pull a top-three from.

What you'll get

You open your inbox in the morning — and instead of two hundred emails, three: "the meeting moved to 3pm", "invoice from the contractor, pay today", "Peter replied about the project." Each with one line on why it's up top. The rest is folded away, waiting for when you have time. You read the three that matter in a minute, instead of scrolling past two hundred.

Weekend plan

1. Saturday. Export 20 of your emails into a plain file. Run them through Luna with the prompt above, get a {score, reason} for each, sort by score. You'll already see that filtering by meaning beats words.
2. Sunday. Hook up a real stream — an inbox export, an RSS feed, or a channel. Have the script hand back the top three, with reasons, once a day.
3. Point it at whatever actually drowns you — email, job posts, a chat — and keep only what matters.

Start with 20 emails in a file. Leave "a filter for every inbox on earth" for later — first let the model reliably pull three that matter out of twenty.

Источник: GPT-5.6: Luna at $1/$6 — a cheap tier for volume (UsageBox)