Zhao Yang’s team did something not very common among startups: it paid people to use its product, 20 to 30 US dollars each. The product was AI live streaming; the team worked on it for half a year, had been launched for three months, and accumulated about 200 users. When payment stopped, the people stopped coming (interview compilation[1]).
In the same account there is a detail more worth looking at: the only thing those 200 people liked was the editing feature the team added last to the live-streaming product—press a button during a live stream to mark a moment, and after the stream ends it automatically produces edited clips. The team spun it out into a standalone product and named it Opus Clip; within two weeks, tens of thousands of users signed up.
What OpusClip does can be explained in one sentence: it cuts a long video into many pieces that can be posted on short-video platforms. For example, put in a podcast or interview, and it picks out several segments, crops them vertical, adds subtitles in more than 20 languages that follow the spoken content, gives each clip a “Virality Score,” and then with one click sends it to YouTube Shorts, TikTok, or IG Reels, or downloads it. On the pricing page, this whole set is placed in the Starter tier (OpusClip pricing page[2], retrieved 2026-11-14).
Image: source OpusClip actual image source[3], retrieved 2026-11-10, original interface retained.
It takes over the most annoying part, not the part requiring the most judgment
In the past, making a clip meant dragging through the entire long video from beginning to end, remembering which lines sounded like highlights, then going back to trim the ends, crop the horizontal frame to 9:16—the person is not necessarily in the middle of the frame, so you have to reframe; then check the subtitles line by line, and finally export, rename, post to platforms, and write a title. None of these steps is difficult; added together, they are a stretch of tedious manual work.
OpusClip takes over the mechanical work in the middle: finding segments, cutting, reframing, subtitling, laying out, and scoring. It hands over candidate clips and scores together, but what to record, which sentence is worth pulling out on its own, and whether the clips the AI picks are usable still require a person to select and revise. In the interview, Zhao Yang described this type of product as AI doing 80% and humans filling in 20% (interview[1])—that is a product-design tradeoff, not a success rate.
This also explains why the team was willing to pull that editing feature out on its own. The editing feature only handles things that have already been shot; what it replaces is repetitive labor. In the interview, Zhao Yang said the live-streaming tool “did not help them save time, nor did it really help them make more money”; on the OpusClip side, he attributed users proactively reaching out—even calling—and being willing to pay to “we did indeed save them dozens of hours.”
One feature stayed, and the team moved on to a standalone product
The company was founded in January 2022; for half a year, the team put everything into AI live streaming, and what it got in return was three months that required paying people to use it. Those 200 users had been paid to come, so their feedback should not have been treated as market validation in the first place. But one action was very concrete: they asked, “Can we use only this feature?” After the money stopped, the people stopped coming, and the only thing named as something to keep was also this feature. The same batch of data can be read as “the product needs improvement” or as “the product choice was wrong.” Judging by actions, he bet on the latter.
Between $15 and $29, the difference is not just quota
Spread the prices out, and the business design is more direct than the feature table.
Free clips cannot be exported after 3 days. This “three days” is the hardest hook in the whole funnel: you edit first, and when you finish, you find that the thing cannot be kept.
Starter is $15 per month, with 150 credits per month. Pro is $29, or $174 per year, which the page converts to $14.5 per month (pricing page[2], retrieved 2026-11-14).
What makes me frown is storage and export: Starter’s media storage expires after 29 days, and MP4 export is limited to 30 days; only Pro’s MP4 export has no time limit, and the storage column gives 100GB fixed storage without stating an expiry time. For people making content, these two deadlines pinch in the wrong place: a clip might be dug out and reposted three months later, but storage and export cannot wait that long; whether it can be recovered after expiry is not explained on the page, and whether it can be recovered after upgrading is also unclear. This looks more like pricing “archiving” separately than pricing compute.
The line between API and import also sits above Pro: Pro opens only limited API access, while Business opens the full API and custom integrations, as well as automatic import from a CMS or other cloud servers (pricing page[2]). This line shows that the customers it wants to serve are moving from “individual editing” toward “teams with content libraries,” but the public pricing page does not clearly state the read/write permissions and data residency of these interfaces; anyone who wants to connect a CMS has to ask.
Subscription plus credits is not unique to OpusClip either; Submagic’s help page states that credits are deducted on upload and export is free (Submagic pricing[4], credits explanation[5], page updated 2026-10-25). This only shows another company’s rules; how OpusClip’s credits are deducted item by item, and how overages and refunds are calculated, is not written on public pages, and Submagic’s rules cannot be applied to it.
Demand is real; profit is still invisible
What shows demand more is what the product looked like on day one: when OpusClip launched, it had no payment function at all, but quite a few users proactively sought it out, even called, saying they were willing to pay to jump the queue, wanted higher resolution, and wanted more downloads (interview[1]). Users proactively asking to pay is more of a demand signal than any conversion rate, but it is still only a signal.
The interview gives revenue numbers: by the end of 2023, seven months after the product launched, ARR was close to $10 million, and registered users were 5 million (interview[1]). This is the founder’s own account, not an audited financial report; it does not include the number of paying customers, MRR, refunds, or gross profit, nor does it give customer acquisition cost, channel mix, or retention distribution. The pricing page has entry points for a free trial and a free plan, but how many free users become paying users is not in the materials.
The cost side is likewise blank: compute, storage, licensing, support, and personnel are all undisclosed, so there is no basis for inferring profit from plan prices or ARR. What can be established is that this is a business that sells processing capacity by credits; billing by credits means that the more usage there is, the more likely compute costs are to rise with it, but the specific bills are not disclosed.
Zhao Yang treats “willingness to pay out of one’s own pocket” as the dividing line for PMF. By that line, those 200 people did not give a positive answer: they were using it with money provided to them. What comes closer to an answer is what happened later—users proactively reached out, even called, willing to pay to jump the queue, wanting higher resolution, wanting more downloads; Zhao Yang attributed this to the product indeed saving them dozens of hours. Acknowledging that those 200 people did not give a positive answer is harder than adding another feature.