Back, and this time it's weekly

The cheapest excuse in your business just got cheaper

What happened: On July 30 OpenAI cut GPT-5.6 Luna by 80%, from $1 to $0.20 per million input tokens and from $6 to $1.20 output. Terra fell 20%, and flagship pricing did not move. Source

1. ChatGPT Shopping opens to every US user

   Why it matters: Etsy and a million-plus Shopify merchants are connected, and Google shipped agentic checkout in AI Mode. If your product data is messy, you're invisible to buyers who never see your storefront.

   DROPIFIED · 4 min read

2. The models are optimizing themselves now

   Why it matters: OpenAI says its own Sol model rewrote the GPU code behind these gains. The price floor keeps falling on a schedule nobody controls, so the tool you rejected on cost last year deserves a fresh look this quarter.

   THE RUNDOWN · 3 min read

A market monitor that watches five sources while you work

This is the research role of my own system, cut down to the one job worth building first. If I lost the whole setup tomorrow, this is the piece I'd rebuild first.

It's for any owner who keeps meaning to check five things and checks them twice a year instead. One sitting to set up, then it runs without you.

1. List five sources. Five, not fifteen. Only the ones where a change would actually alter a decision you make.
2. Write the change definition. "Competitor pricing" is useless to an agent. "Their starting tier moves by more than 10%" is a trigger it can act on.
3. Paste the prompt. Open a new project in whichever assistant you already pay for, drop the prompt below in as the instructions, and fill the brackets from steps 1 and 2.
4. Schedule it. Hourly during business hours covers most businesses. Start slower than you think you need to.
5. Keep the silence rule. An agent that reports "nothing happened" every hour trains you to ignore it, and then you miss the one that mattered.
6. Read a full week first. A wrong alert gets fixed in the trigger definition, never in the output. That's the difference between training an agent and babysitting one.

You are the market monitor for [BUSINESS NAME]. We sell [WHAT YOU SELL].

Your job is to check these sources on every run and report only what changed:

1. [SOURCE 1] / report only if: [TRIGGER 1]
2. [SOURCE 2] / report only if: [TRIGGER 2]
3. [SOURCE 3] / report only if: [TRIGGER 3]
4. [SOURCE 4] / report only if: [TRIGGER 4]
5. [SOURCE 5] / report only if: [TRIGGER 5]

On every run:
- Compare what you find against what you found last run. Keep a short record of each source's last known state, so you can tell a real change from a page that reworded itself.
- If nothing crossed a trigger, output exactly: NO CHANGE. Send nothing to [YOUR EMAIL].
- If something crossed a trigger, send one message to [YOUR EMAIL] with this structure:

  What changed: one sentence, with the number or the quote in it.
  Source: the direct link.
  Why it matters to us: one sentence, in terms of our revenue or our costs.
  What I would do: one action, or "watch only" if nothing is warranted yet.

Rules:
- Never report a change you cannot link to. If you cannot verify it, say so instead of guessing.
- Never send more than three items in one message. If more than three triggers fire, send the three with the largest financial impact and say how many you held back.
- No summaries of things that did not change. No preamble, no sign-off.
- If a source is unreachable, name it and say for how long. A silent failure is worse than a missed alert.

Walk away with: a monitoring agent that watches five sources all day and only speaks when something actually changed

The real return here isn't the line item, it's the attention. Checking is the kind of work that never finishes and never shows up on an invoice, so it gets postponed until something has already moved. Hand it to something that doesn't get bored and you stop carrying it around in your head.

Build that one and the second takes half as long, because you already know what a trigger definition looks like. That compounding is the reason to start with something small enough to finish this week, and it's the reason I spent months building the Agentic AI Community. One agent is a start. A workforce is a business.

A price cut is the easiest thing in this industry to get excited about, and the least likely to change your year.

Here's what I see. Owners spend a weekend comparing benchmarks, pick a winner, open a chat, and type the same vague request they typed into the last one. The model changed. The instructions did not. So the output doesn't either. My setup runs across three models and switching takes about a minute. The part that took real work was writing down how my business makes decisions.

Where I might be wrong: if a model showed up that produced work matching my standards without being told what those standards are, I'd drop this position tomorrow. Every release gets closer, and none of them are there yet.

Best,
Taki Wong
@taki.gpt