OpenAI Posts Video of Perplexity Testing Code With Astra

In the video, which METAL reviewed in full, Ho introduces himself as a Perplexity co-founder and says he has worked on product and engineering for the past four years. OpenAI's page lists his title as co-founder and chief strategy officer. He explains that Perplexity has always been centered on search and accuracy, and that the ability to process large amounts of information has been very important to the company.

The first claim in the video is about the link between code and search. "Every single time the model gets better at writing code, search and accuracy improve," Ho says. The reason, he explains, is that the model writes the code that searches the web or internal information and then combines the results into a concise summary a person can read. For an answer engine, coding ability is not a side project; it is the lever that lifts search quality.

Then comes the heart of the video. According to OpenAI's page, Ho has limited time to test manually, so he asks GPT-6 Astra to build a small testing program around an application. The model generates realistic responses like those another service would send, such as a language model API or a connector, and stands in for those services. That lets him check how the application responds and test the workflow from start to finish.

Seen through an engineer's eyes, this reverses the usual order of test automation. Normally a person writes fake responses ahead of time and runs tests on top of them. Here the model takes over the job of producing the fake responses too. Since the model already knows what an external service is likely to return, imitating it is handed to the model as well. It is a bit like a carpenter building the formwork they will pour into themselves.

OpenAI's page lists three jobs Perplexity hands to the model: drafting communications, changing software, and monitoring production systems. In the video, Ho says these are things earlier generations of models could not do. He frames the challenge that came before it as taking the informational side of Perplexity and applying it to real-world systems as AI keeps improving.

The sentence that follows is the conclusion of the case. "We're actually able to trust it with full end-to-end systems and check in on it much less frequently than previous generations of models," Ho says. In the video he corrects himself mid-sentence, from "less frequently" to "much less frequently." The page's summary line says the same thing: Perplexity checks in far less often than it did with earlier models. He closes the video by saying the company wants to deliver that capability to customers in new form factors as well.

The video is one of a run of customer stories OpenAI is releasing around GPT-6 Astra. METAL reported earlier on Cognition's Devin proving its own work with the same model, and that story also covered the first quote from the Perplexity case page. This time a founder speaking on camera has been added to what was only a written case. METAL also covered OpenAI opening the Codex harness through its API. Opening the harness outward, then putting forward a customer who runs its checks on that harness, is one continuous line.

The details shown on OpenAI's page are modest. Company size: startup. Region: North America. Industry: technology. Product used: API. Within a few hours of posting, the video passed 38,000 views and drew 73 replies. The picture OpenAI wants to show is clear. The model does not stop at writing code; it builds the apparatus that checks whether the code is right, and people look at the results less often. Perplexity says it is already working that way in production.