A Chinese food delivery company just released one of the most capable coding AI models on the planet. And it did so at a fraction of what OpenAI and Anthropic charge for their flagship products.

Meituan, the Hong Kong-listed on-demand services giant best known for delivering dumplings and booking hotel rooms, unveiled LongCat-2.0 on June 30. The model scores 59.5 on SWE-bench Pro, a widely watched benchmark for evaluating AI performance on real-world software engineering tasks. That edges out GPT-5.5’s score of 58.6 and surpasses Claude Opus variants on the same test.

Here’s where it gets interesting for anyone building with AI: the API pricing. LongCat-2.0 comes in at roughly $0.75 per million input tokens and $2.95 per million output tokens. If you’ve been watching what Western frontier models charge for comparable capability, those numbers look almost confrontational.

Built without a single Nvidia chip

LongCat-2.0 is a Mixture-of-Experts (MoE) architecture. In English: instead of running every calculation through the entire model, it routes each request to specialized sub-networks, activating only the parameters it needs.

The total parameter count sits at 1.6 trillion, but the model dynamically activates between 33 billion and 56 billion parameters per token, averaging around 48 billion. The model also ships with a native 1-million-token context window. That’s enough to process entire codebases, lengthy legal documents, or multi-chapter manuscripts in a single pass.

Perhaps the most geopolitically loaded detail: LongCat-2.0 was trained entirely on a cluster of 50,000 domestic Chinese chips. No Nvidia A100s. No H100s. Meituan built this without access to the cutting-edge American hardware that US export restrictions have explicitly tried to keep out of Chinese AI labs.

Open source, open season

Meituan isn’t keeping this behind a paywall. The company open-sourced LongCat-2.0 under a permissive MIT-style license, with model weights available on both GitHub and Hugging Face.

Before the official reveal, the model had already been turning heads. It reportedly topped global developer leaderboards under the alias “Owl Alpha,” a quiet flex that let the model’s performance speak before its corporate parentage was known.

This follows a playbook that Meta pioneered with Llama and that DeepSeek has executed with striking effectiveness in China. The difference here is that Meituan isn’t primarily an AI lab. It’s a services company with 700 million annual transacting users, and it’s already integrating earlier LongCat models into its consumer-facing products for restaurant recommendations, hotel suggestions, and merchant tools.

The model’s design emphasis on agentic coding tasks, including code understanding, generation, execution, and multi-step reasoning, positions it squarely in the enterprise AI market where margins are fattest and switching costs are highest.

What this means for the AI market and investors

When a model that benchmarks competitively with GPT-5.5 comes in at roughly $0.75 per million input tokens, it puts direct pressure on the pricing power of every Western AI provider.

There’s a geopolitical dimension investors should watch carefully. LongCat-2.0’s training on entirely domestic Chinese hardware demonstrates that US export controls haven’t created the capability gap they were designed to produce, at least not at the timeline policymakers expected.

The risk side is worth noting too. Meituan trades on the Hong Kong Stock Exchange under ticker 3690.HK, which means it operates under Chinese regulatory oversight. Any enterprise or developer building critical infrastructure on LongCat-2.0 inherits that jurisdictional exposure, a consideration that cuts both ways depending on your risk tolerance and where your users are located.