A 550B Model Just Sat in Your Focus Group. It Sounded Like a Buyer.

This week a lab shipped a 550-billion-parameter model whose only job is to pretend to be your customer.

Not write ads. Not draft a deck. Sit in a multi-person conversation and drip information the way a real buyer does.

humans& called it Persimmon. On a multi-user Turing-style test it fooled LLM judges 18 to 21 percent of the time. Frontier assistants sit under 3 percent. Same window, OpenAI put GPT-Live-1 in the API at five cents a minute and wired PitchBook, LSEG, and Crunchbase into ChatGPT for Financial Services.

If you run a marketing agency, that is not a research paper. That is a cheaper focus group, a voice channel, and a junior analyst that does not sleep.

What actually shipped this week

LAYER 1 — Synthetic buyers got usable

Persimmon v0.1 is a 550B user model built on NVIDIA Nemotron 3 Ultra. It is trained to simulate open-ended, multi-user talk, not to be a helpful assistant. It keeps long-horizon coherence across 80 turns and times information the way people do. That is a panel you can rerun at 2am without booking a room in Gurgaon.

LAYER 2 — Voice stopped being a demo

OpenAI opened GPT-Live-1 to developers. Full duplex. It listens and speaks at once. Reported turn-taking latency is about 0.8 seconds. Price: $0.05 per minute for the voice layer. WhatsApp already let third-party agents into the thread. Now the voice stack can sit on top of it.

LAYER 3 — The analyst seat got productized

ChatGPT for Financial Services launched on GPT-6 Astra with built-in market data and design partners at Morgan Stanley and Evercore. Clay raised another $115 million to keep building GTM agents. Amazon plugged DSP inventory into ChatGPT Ads. The conversation is now a media buy and a research desk.

The marketing shift nobody should ignore

Most teams will read those headlines as ‘cool tools.’ That is how you stay a production house while the client starts testing offers without you.

AI for marketing agencies in 2026 is three jobs on one desk:

• A synthetic audience that pressure-tests the offer, the landing page, and the first three objections before you spend media.

• A voice or chat agent that can hold the conversation after the click — not a FAQ bot, an owner of the thread.

• Answer-engine presence so ChatGPT and Gemini can quote your price and proof without inventing them.

We are running it that way at Dzine Prodigy. One live brief. Five synthetic buyers. One agent on research and first draft. A human signs the version that survives. No transformation offsite. No 40-tool stack.

Do this next week, not next quarter: pick a campaign that is about to spend. Run the offer through a simulated buyer panel. Fix the sentence that dies in every conversation. Then buy the media.

Labs will keep shipping models. Agencies that treat this week as a newsletter will still be collecting logos in October. Agencies that ship one tested offer will have a case study.