The best AI launch videos on X, and why they worked
By Laza · October 16, 2026
Most AI companies launch with a video now, and most of those videos land on X first. A handful became reference points for everyone who launched after them. Here are six of them, what the video actually did, what the reporting says happened next, and the lesson we take from each.
None of these were made by us. Every post below is public; click through and watch the originals. Our launch video gallery has stills, lengths and hook lines for each one.
Devin by Cognition: claim the category, then get checked
The Devin post (March 12, 2024) opened with “the first AI software engineer” and a video of Devin working inside its own workspace. Cognition’s launch post backed the claim with a benchmark: Devin resolved 13.86% of issues end to end on a subset of SWE-bench, against a previous best of 1.96%. The post also said the team had given Devin real jobs on Upwork.
The launch worked: it set the frame for every AI coding agent that followed. It also drew scrutiny. A month later The Pragmatic Engineer covered an engineer’s analysis concluding that Devin did not complete the Upwork task the way the video implied, and reported Cognition saying Devin “definitely makes mistakes or often fails” on such jobs.
Lesson: a category claim gets attention and an audit. Every claim in your launch video has to survive someone rewatching it frame by frame.
Sora by OpenAI: the output is the pitch
OpenAI’s Sora post (February 15, 2024) was a 17 second clip with the prompt in the text. No interface, no narrator. MIT Technology Review reported that Sora could produce clips up to a minute long and that OpenAI had no plans for a public release yet, and pointed out that the samples were likely cherry picked, with flaws visible in some of them.
Lesson: when your product makes something visual, the output is the launch video. Put the input (the prompt) next to it so viewers can judge the gap themselves.
Figure 01 with OpenAI: answer the doubt in the first frame
Figure’s post (March 13, 2024) opened on a title card that included “film speed 1.0x”. The first question anyone asks about a robot video is whether it is sped up or remote controlled. MobileSyrup quoted founder Brett Adcock saying there was no teleop and that the video was filmed at 1.0x speed and shot continuously, and quoted a Figure engineer saying all behaviors were learned, not teleoperated.
Lesson: find the first skeptical reply your video will get and answer it on screen before anyone posts it.
GPT-4o by OpenAI: one live interaction beats ten features
The GPT-4o post (May 13, 2024) showed one person talking to a phone in real time. CNBC reported that the model was announced in a livestreamed event with live demos, including real time translation between Italian and English, and that it could respond to audio in as little as 232 milliseconds.
Lesson: if speed is the feature, show it at real speed with a human in the frame. A spec sheet number means less than watching the pause disappear.
bolt.new by StackBlitz: one post, no campaign
The bolt.new post (October 3, 2024) asked a question that named tools developers already used, then showed a 43 second run from an empty prompt to a working app. Contrary Research writes that the team shipped it without any marketing, “just a single tweet”, and describes the launch as true virality across channels.
Lesson: start the video at the blank state your buyer knows, and get to the result fast. A short, clear demo can carry a launch on its own.
rabbit r1: long form works when the event is the news
rabbit’s post (January 9, 2024) was a full keynote, over 25 minutes, posted natively during CES with an order link. BGR reported about 10,000 preorders on the first day and quoted rabbit saying it had hoped to sell 500 devices on launch day and beat that by 20x.
Lesson: a long video can work on X when the event itself is the reason to watch. Most launches are not events, so plan short cuts for the rest of the week.
What they share
- One sentence. Each launch can be summed up in a line, and the video proves that line.
- Product on screen early. Output, interface or the robot itself, within the first seconds.
- A built in answer to doubt. Real speed, the prompt in the text, a live human.
- Claims that hold up. Devin shows the upside and the cost of a bold claim.
Want the annotated set offline, with a checklist for your own cut? Get the swipe file. Ready to make yours? Send us your brief.