OpenAI published a customer story on October 8, 2026 about Pollo AI, an Asia-Pacific platform that turns a prompt or a product photo into finished images and short video ads. The page is not a product launch. It is OpenAI using one of its customers to describe, on the record, which of its models a 26-million-user creative product runs in production, and which job each one is given.
What was actually released
Nothing new ships on OpenAI’s side. What is disclosed is a stack. Pollo AI runs three OpenAI models, each on a separate job.
| Job | Model | What the page says |
|---|---|---|
| Routing and lighter requests | GPT-5.6 | Handles the first pass and decides how much reasoning a request needs |
| Narrative and scene planning | GPT-6 Astra | Used for storylines, scripts and scene revisions, the steps the page calls the most challenging |
| Image generation | GPT-Image-2.5 | Used for every image job: posters, storyboards and campaign visuals |
Pollo AI says the split exists to match reasoning depth to the task, so a light request is not answered by the most expensive model. The company’s latest offering is Pollo Agent, a guided workflow that takes a rough idea and a visual reference and develops a storyline, a scene flow and a script before it generates anything. Pollo AI’s Marketing Studio also carries a Photo to Video Ads tool that turns a still product image plus a prompt into a short clip: the page walks through a fragrance ad of about 15 seconds built that way.
Scale is given in passing: more than 26 million users, and 30% of creator sessions starting from a template or a trending-format workflow rather than from a blank prompt.
What the claim rests on
Every number on the page is a company figure, repeated by OpenAI without a method. Four of them carry the piece.
| Claim | Figure | Basis published |
|---|---|---|
| Users | more than 26 million | none, and no split between free and paying |
| Sessions starting from a template | 30% | none; described as company analytics |
| Time spent choosing or switching models | more than 50% less | none; no baseline, no sample, no window |
| Length of the ad in the example | about 15 seconds | one customer example, not a range |
The 50% figure is the one a buyer will repeat, and it is the least checkable. Less than 50% compared with what: an earlier version of the product, a manual video workflow, another tool? The page does not say, and it publishes no sample and no measurement window. “Pollo AI reports” is the only attribution given.
The same limit applies to the model claims. A founder saying GPT-5.6 performed especially well is a testimonial, not a measurement, and a customer saying every image job runs through GPT-Image-2.5 is a product decision, not a test result. No latency, no cost per generation, no error rate and no head-to-head output comparison appear anywhere on the page.
What is not published
No price, no plan tiers, no rate limits, no data retention terms and no independent test of the images or clips produced. A case study is the wrong place to expect them, which is exactly why nothing here should be read as a benchmark.
On this page
What it changes for someone buying a creative tool
Three things are worth taking from the page.
First, routing. The production pattern is now explicit: a lighter model for the first pass, a stronger one for the reasoning-heavy step, and a dedicated image model for visuals. If you pay premium rates for a tool that sends every request to one top-tier model, part of that bill is being spent on work a smaller model could do. When you compare tools, the useful question is which model handles which step, not which model the vendor puts on the marketing page.
Second, text inside images. The CTO’s line about characters and letters names a real failure mode of ad creative: posters and storyboards are useless when the lettering comes out garbled. It is a specific enough claim to test on your own brief before you commit.
Third, short-form production. The promise is a product ad in minutes from a still image, and the page shows one example of about 15 seconds. Treat it as a direction, not a guarantee, and run it on your own product photos rather than on the demo.
Ask a vendor three questions: does the tool route per task, which models does it call, and what happens to your images after generation.
The Neural Minute take
The number worth keeping is not 26 million. It is three.
A platform with that many users does not settle on one model and stop. It runs a split, and it is the split that makes the page worth reading: routing on a cheap model, reasoning on a strong one, images on a specialized one. That is the same shape we described when Google Cloud put hard spend caps on its Gemini agent, and it keeps showing up wherever a creative tool has to hold a cost line.
We want to do for video what Canva did for design.
The analogy is more useful than the growth numbers. Canva did not win by being the most capable design tool; it won by removing the expertise barrier, and that is the bet here: not better clips, but a shorter path from an idea to something postable. The user counts are a sign that the path works, not proof that the output beats what a competitor produces.
Our reservation is with the 50%. It is the figure a buyer will quote internally to justify a switch, and it is the one the page gives no way to check. For a claim about saved time, the arithmetic is cheap to publish and would settle the question: a baseline, a sample, a window. None of the three appears. Treat the model split as a signal about how serious production tools are built, and treat the savings figure as a number to ask for in writing, the way you would with any supplier quoting a percentage.
Sources
- Pollo AI turns creative ideas into campaigns with OpenAI, openai.com, published October 8, 2026: https://openai.com/index/pollo-ai
- The three-model split, the 26 million user figure, the 30% template share, the 50% time claim and the 15-second ad example are read from that page and reproduced here as company statements, not as independent measurements.
- No price, benchmark or third-party test of Pollo AI appears on the page, so none is quoted here.