Banks Just Tripled Spend Inside ChatGPT. Your Category Is Next.

Last week was the $1 billion headline.

This week is the mix shift. And that is the number agencies should actually care about.

Sensor Tower just published what is sitting inside ChatGPT Ads. Financial services went from 2 percent of spend in April to 13 percent in August. Shopping and retail gave up the room. Software, travel, and health followed the banks in.

Same four days, OpenAI put GPT-6 Astra into paid ChatGPT and called it the AGI era. The New York Times and the Seattle Times kept stacking copyright filings. Corporate buyers quietly moved more work onto cheap open models.

If you still treat AI as a copy assistant, you are measuring the wrong week.

What actually shipped

I am not collecting headlines for a newsletter. I am watching the media plan change under the campaign calendar.

ChatGPT Ads is no longer a test format. It is a $1 billion run-rate channel with self-serve live in India, Europe, the Middle East, and North Africa. The new fact is who is buying it. Banks do not experiment with vanity inventory. They buy conversations that sit next to intent.

Astra is not a caption model. It is a computer-use model. Spreadsheets. Browsers. Multi-step professional work. That is the production layer sitting behind the same product the ads now live in.

Qwen3.8-Max just tied the frontier on Code Arena at about $5 per million tokens. AT&T-class buyers are already routing work to open weights. The closed stack is still useful. It is no longer the only seat in the room.

The marketing brief

This is the part most teams will skip, then pay for in Q4.

LAYER 1 — Treat ChatGPT as media, not a writing tool. If financial services can 6x their share in four months, your category can move in a quarter. Build a test: three offers, one proof point each, live in Ads Manager. Measure new-to-brand conversations, not last-click vanity.

LAYER 2 — Make the answer engine quote you correctly. Generative engine optimization is not a blog tag. It is structured facts an agent can lift without inventing your price, your market, or your proof. Clean service pages. Named outcomes. Dates. Numbers a model can defend.

LAYER 3 — Split the stack. Use Astra or Claude when the job is computer use and long work. Use the cheap open models when the job is volume. Do not invoice the client for frontier tokens on a caption pass.

BEFORE: a content calendar and a media plan that never touch. AFTER: one brief that ships the page the answer engine can cite, the ad that sits inside the chat, and the agent workflow that produces the next 20 variants by Friday.

Do this next

Pick one offer. Write the three sentences you want ChatGPT to say about it. Put those sentences on a page an agent can read. Then buy the conversation around the same query the page answers.

That is the whole job this week. Not another model bake-off. A channel, a citation, and a stack that actually ships.