Chinese tech and delivery giant Meituan unveiled LongCat-2.0 on June 30, 2026, simultaneously disclosing that the model had been publicly available for roughly two months on the AI platform OpenRouter under the anonymous name “Owl Alpha” — without revealing its identity or origin.
During that period, Owl Alpha ranked first on OpenRouter’s Hermes Agent workspace, second on Claude Code, and third across OpenClaw deployments, all by monthly call volume, according to Meituan’s announcement. The disclosure suggests that developers worldwide were already relying on a Chinese open-weight model without knowing it.
Architecture and scale
LongCat-2.0 is a Mixture-of-Experts model with 1.6 trillion total parameters, approximately 48 billion of which are activated per inference. It supports a native one-million-token context window — roughly the equivalent of 750,000 words held in memory simultaneously — and was trained on over 35 trillion tokens.
The model was built entirely on a cluster of more than 50,000 domestic Chinese accelerators, with no Nvidia hardware at any stage of training or inference. It is released under an MIT license, allowing free commercial use and modification.
Benchmark performance
On SWE-bench Pro, a standard benchmark for autonomous GitHub issue resolution, LongCat-2.0 scored 59.5, surpassing GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2). On Terminal-Bench 2.1, it scored 70.8. On FORTE, which tests office-task automation agents, it reached 73.2 — matching Claude Opus 4.6 though falling below GPT-5.5’s 77.8. Leading Claude models still hold an edge in overall rankings.
Pricing
The public API is priced at $0.75 per million input tokens and $2.95 per million output tokens — roughly one-sixth the cost of GPT-5.5 ($5/$30) and on par with DeepSeek V4-Pro. Model weights are scheduled to be released on Hugging Face under the same MIT license.