Anatomy of an AI hype cycle — how Devin, Sora and Rabbit R1 traced three different curves
Three AI products launched to massive hype between 2024 and 2025. One quietly compounded its way into Goldman Sachs. One died losing $15M a day. One is dying slowly. The differences were visible on day one.
On March 12, 2024, Cognition Labs posted a video on Twitter titled "Introducing Devin, the first AI software engineer." Within forty-eight hours it had been viewed tens of millions of times. Aravind Srinivas, the CEO of Perplexity, said Devin "seemed to be the first demo of any agent, leave alone coding, that crosses the threshold" of human capability. Cognition closed a Series A at a $2 billion valuation almost immediately.
Three weeks after that launch video, a YouTube channel called Internet of Bugs posted a thirty-minute video taking the demo apart piece by piece. The reviewer showed that Devin had been credited with finishing tasks it hadn't actually finished. That the agent had invented work the original task didn't ask for. That the timeline shown in the demo had been compressed to make it look much faster than it was. The video racked up over a million views.
It looked, at that point in early 2024, like the moment Devin died. Two years later, the company is embedded inside Goldman Sachs across a 12,000-person engineering organisation. Cognition has merged hundreds of thousands of pull requests across thousands of customer companies. Annual recurring revenue more than doubled in the eighteen months following the backlash, and the company that looked dead in April 2024 has quietly become one of the most successful enterprise AI deployments in software.
What actually happened to Devin between then and now isn't the curve most people watching the launch predicted. Two other high-profile AI products from the same era ended up on more familiar trajectories, and reading their stories alongside the Devin one gives a rough framework for spotting which products will compound and which will collapse, often well before the data is fully in.
Most AI launches in the current cycle, after the opening spike of attention burns off, end up settling onto one of three rough trajectories.
Sharp launch peak, backlash within weeks, decline ending in shutdown or zombification. The economics or the category don't survive contact with reality. (Sora.)
Modest launch peak that doesn't quite collapse but never recovers either. Diminishing user numbers, increasingly desperate pivots, eventually irrelevance. (Rabbit R1.)
Launch peak, brutal backlash that looks like collapse, quiet pivot to a different audience, then growth that compounds out of public view. (Devin.)
The differences between these curves don't really announce themselves at launch, because the launch part of any AI product story tends to look fairly identical from the outside: a viral demo on Twitter, strong opening numbers, a founder media tour, and follow-on funding within weeks. The divergence between the curves only becomes visible somewhere between week six and month four, and what drives it tends to be structural features of the product that were knowable on day one if anyone was looking at the right things.
Sora
OpenAI · Standalone consumer video generation
Sora launched into a category that combined two characteristics that should probably have been read as warnings before the product ever shipped. Inference costs were significant: each ten-second clip required roughly forty minutes of parallel GPU processing and cost about $1.30 in compute. Willingness-to-pay was always going to be modest, since most consumers using video generation were doing it for entertainment or curiosity rather than for revenue-generating work. The unit economics were going to be difficult from the start.
The launch itself worked. Number one on the US App Store for twenty consecutive days. Faster to a million users than ChatGPT itself. Downloads peaked at 3.33 million in November 2025. A planned Disney partnership covering more than 200 characters from Marvel, Pixar, and Star Wars was announced in December.
Then the math caught up. Forbes, citing analyst estimates from Cantor Fitzgerald, reported Sora was burning roughly $15 million per day in inference costs at peak. Mobile intelligence firm Appfigures put Sora's lifetime in-app revenue at $2.1 million. By February 2026, downloads had fallen 66% to 1.13 million. The Disney partnership evaporated. On March 24, 2026, OpenAI announced the shutdown.
What was visible at launch. The unit economics. A consumer AI product whose per-use compute cost runs above roughly 10% of its monthly subscription price tends to be in the same equation Sora was in, and that equation generally doesn't get easier as users scale; it gets worse.
Rabbit R1
Rabbit Inc · Standalone AI hardware · "Large Action Model"
Rabbit's launch at CES 2024 was the high point of the AI hardware moment. The company sold roughly 100,000 units of a $199 device that promised an autonomous "Large Action Model" capable of taking actions across the apps on your phone. Pre-orders sold out. Tech press coverage was breathless.
Then the units shipped. Reviewers found voice-response latency of up to ten seconds. The "Large Action Model" couldn't reliably interact with most third-party apps. The device had four functioning integrations: Spotify, Midjourney, DoorDash, and Uber. Basic phone tasks (timer, email) failed. Battery life was about an hour. Engadget called it "a $199 AI toy that fails at almost everything." Wired, Gizmodo, and Inc. published variations on the same finding.
The R1 didn't die on impact, though. It limped. RabbitOS 2, released in September 2025, redesigned the interface and rebranded the device as an "AI agent assistant" rather than the original autonomous-agent concept. By early 2026, reports of unpaid employee salaries and financial distress had begun to surface. Jony Ive publicly criticised both the R1 and the comparable Humane Pin, and the entire category of standalone AI hardware appears to have been a category error in retrospect.
What was visible at launch. The category position. Both the R1 and the Humane Pin were effectively selling against the iPhone, which was always going to win that fight given Apple's distribution and Apple's own AI roadmap. The useful question to ask of any AI product at launch is probably less about whether the product is good in isolation, and more about what it looks like once the dominant adjacent platform ships the same capability as a feature.
Devin
Cognition Labs · Autonomous AI software engineer
Devin's path is the most useful of the three to study, because unlike the other two, the original product more or less worked. It just didn't work as well as the marketing implied at launch, and the gap between the demo and the reality became the dominant story for several months after.
The honest read of the original Devin, available now that two years have passed, is that it scored 13.86% on SWE-bench Lite. That figure was a genuine improvement over prior state of the art at the time, but it also meant the product was failing on more than 86% of tasks. By early 2026, multiple competing systems were scoring above 75% on the much harder SWE-bench Verified, and Devin's lead on benchmarks had effectively evaporated.
What happened instead is that Cognition pivoted in three directions that compounded. They acquired Windsurf, the AI-native IDE, which gave them a real-time coding surface to complement Devin's autonomous batch surface. They built Devin Review, a code review tool that addressed what Cognition's customers were telling them was now the bottleneck — that AI agents were generating more code than humans could review. And they went hard into enterprise. By July 2025, Goldman Sachs had begun deployment to its 12,000-person engineering organisation, with hundreds of Devin instances running and a path to thousands.
The metrics from inside Cognition tell the rest of the story. By early 2026, Devin sessions per week across all enterprise customers had doubled in just six weeks. In their best week of 2025, the team merged 154 internal Devin-generated PRs. By early 2026 that figure was 659 in a single week. PR merge rate across enterprise customers improved from 34% to 67% year-over-year.
The hype cycle hit Cognition harder than most because they fed it harder. But the underlying business they were building was a B2B platform play, and B2B platform plays don't generally die from a YouTube takedown. They die from failing to find customers willing to pay six figures for something that solves a real bottleneck. Cognition found those customers.
The structural feature that was visible at Devin's launch, which most people missed in the noise around the demo, was that Cognition had a strong technical team (10 IOI gold medals on the founding roster), real research depth, and a willingness to operate on enterprise sales timelines rather than consumer ones. Most of Devin's enterprise customers stay quiet about using it. Goldman Sachs is the public exception, and that quietness from the rest of the customer base is itself a useful signal about what's actually happening underneath.
The interesting question is whether you can tell which curve a product is on early enough to act on it. The signals are different for each.
| Job | Default pick | Why |
|---|---|---|
| Collapse curve | Read the unit economics on day one | A consumer AI product whose per-use compute cost exceeds ~10% of its monthly subscription price is in the Sora equation. The equation does not improve at scale. |
| Slow-death curve | Read the category position | If a tool is selling against a dominant adjacent platform that is about to ship this as a feature, the timeline is already running. iPhone, ChatGPT, Gemini, and Workspace are the four big absorbers in 2026. |
| Resurrection curve | Read who is quiet about using it | AI products working in production tend to have customers who aren't loud, because the people deploying them are inside risk-averse enterprises. Watch for credible deployments inside otherwise-silent companies, followed by case studies two or three quarters later. |
The implication for picking AI tools right now is that the tools worth tracking aren't necessarily the loudest ones. They're the ones with second-order signals: slow-burn case studies, migration posts from large companies, named hires from enterprise customers. The hype cycle moves at the speed of the news cycle. The actual market moves much slower, and ends up in a different place.
Sora's collapse generated bigger headlines than Devin's quiet rise probably ever will. The lessons sitting inside the two products are quite different though, and the more useful one for anyone building or picking AI tools right now is probably the quieter one. Compounding usually doesn't make for good copy.
The signals beneath the headlines
The hype cycle is loud; the actual market is quiet. AgentTape tracks AI products by usage flow, retention, sentiment, and migration patterns — the second-order signals that distinguish the resurrection curves from the collapses. The leaderboard for the part of the market that isn't on the news yet.
View the live indexes