Since late April 2026, an anonymous model dubbed "Owl Alpha" had been topping the usage charts on OpenRouter, with no information circulating about its creator. On 30 June, the veil lifts: it's LongCat-2.0, a 1.6-trillion-parameter model built by Meituan, the Chinese app... for food delivery. And it was reportedly trained entirely on Chinese chips, without a single Nvidia processor.
In the small, very chatty world of AI, there is a discreet practice: releasing a model without a name, under a pseudonym, to observe how developers use it before revealing who is behind it. It's a way of testing the product on the real market rather than on lab benchmarks. But usually, this little game of hide-and-seek lasts a few days, two weeks at most. This time, it lasted two months, and the result surprised everyone.
A silent dominance
Since late April 2026, a model named Owl Alpha appeared on OpenRouter (a platform that centralises access to dozens of different AI models), with no information whatsoever about its creator. No press release, no company name, nothing. Yet this ghost model began crushing the usage charts: first place on the Hermes Agent workspace, second on Claude Code, third on OpenClaw — three channels widely used by developers for AI-assisted coding.
The volume figures are dizzying: roughly 10.1 trillion tokens processed per month, an average of 559 billion per day, with 242% growth in a single month. A model with no name, no marketing, no official announcement, becoming one of the most widely used AI coding tools in the world. Rumours started circulating in early June, but official confirmation only came on 30 June 2026.
The reveal: Meituan, the food delivery app
That day, the official LongCat account announced that Owl Alpha was them. And LongCat is an artificial intelligence project from Meituan, a Chinese company known to the general public for its food delivery app, the Chinese equivalent of Uber Eats. The model is officially called LongCat-2.0, and its technical specs are impressive: 1.6 trillion parameters in total, with a Mixture-of-Experts architecture that activates only around 48 billion per request, and a native context window of 1 million tokens.
Meituan claims that LongCat-2.0 was trained and served end-to-end on a cluster of more than 50,000 locally manufactured Chinese chips, without a single Nvidia processor in the loop, neither for training nor for serving users. If verified, this would be the first trillion-parameter-scale model built entirely on domestic Chinese hardware. The training, which used more than 35 trillion tokens, reportedly went ahead "without interruption or irrecoverable loss of stability", according to Meituan — far from a given on hardware that is still little proven at this scale.
What its performance is really worth
On SWE-bench Pro, a test that measures solving real issues from production code repositories, LongCat-2.0 scores 59.5, ahead of GPT-5.5's 58.6. The gap is real but minimal, less than a point, making it more of a statistical tie than a clear win. On Terminal-Bench 2.1, it reaches 70.8. The model still trails Claude Opus 4.7 and 4.8 significantly on several complex scientific reasoning benchmarks, confirming a profile of an agentic coding specialist rather than a universal generalist.
LongCat-2.0's real selling point is price. Its standard rate drops to $0.75 input and $2.95 output per million tokens, with an even lower launch promotion, and above all a policy where re-reading already-cached content costs nothing. That's significantly cheaper than GPT-5.5 or even Claude Sonnet 5 at promotional rates. A boon for developers running agents in loops over long sessions.
What this story really tells us
Beyond the juicy anecdote of a food delivery company building a frontier model, this affair illustrates a deeper shift we have already observed with GLM-5.2: several analysts note that the access restrictions imposed by the US government on its own labs, such as those that delayed GPT-5.6, have paradoxically opened a wide boulevard for cheaper Chinese open-source alternatives that lack these access constraints.
Still, one should keep a healthy distance from the figures being put forward. The benchmarks come from Meituan itself, with no independent verification, and the model's full weights had still not been released at the time of the announcement despite the promise of open access under an MIT licence. What this story proves with certainty, on the other hand, is that a model can now capture a massive share of the global development market before anyone even knows who built it. Performance alone was enough to carve out its place. The name came after.