Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.
On November 2, 2026, LongCat, the AI development team at Meituan, the Chinese food delivery giant, officially released the AI model LongCat-2.5. It follows the preview version "LongCat-2.5-Preview" released on September 25, and combines image understanding with coding ability. The team says it can be used for coding while looking at the screen, design work, 3D tasks and more. It is already available in Meituan's chat service and via API.
Model overview
According to official documentation and media reports, LongCat-2.5's main specifications are as follows.
- Architecture: MoE (a mechanism that switches between multiple expert models) with about 1.6 trillion total parameters, about 48 billion of which are active per token
- Input: Images and text. A design built for images from the start, converting images into its own discrete tokens and handling them together with text
- Context length: 1 million tokens. A single output can be up to 128,000 tokens
- Use cases: Long-running agent work spanning terminals, browsers, screen operation, spreadsheets and design tools
The LongCat team says it improved the model's ability to develop complex codebases and to build projects from start to finish by adding many long task workflows to the training data and running large-scale reinforcement learning.
Developer tool integration and API
The official change log says the model works on top of a variety of AI agent tools, including Claude Code, Codex and Hermes. The API supports both the OpenAI format (/v1/chat/completions) and the Anthropic format (/v1/messages), so existing tools can easily try it just by changing the endpoint.
API pricing per million tokens is as follows. All are limited-time discounted prices, and no end date has been given.
- Input (no cache): $0.30 (CNY 2)
- Input (cached): $0.006 (CNY 0.04)
- Output: $1.20 (CNY 8)
Performance and open weights
According to Unite.AI, in evaluations published by the LongCat team, the model scored 88.4 on Terminal Bench v2.1, which measures terminal work (compared with 70.8 for the previous-generation LongCat-2.0), and 91.9 on BrowseComp, which measures web research (compared with 79.9). Both are figures reported by Meituan itself and have yet to be verified by third parties.
The weights of the previous-generation LongCat-2.0 were released publicly. For LongCat-2.5, the team says it plans to publish the weights on GitHub and Hugging Face, but as of November 3 they have not been released.
CoAI's take
Well-known Chinese AI models include DeepSeek, Qwen, Kimi and GLM, and Meituan's LongCat has now released LongCat-2.5, the successor to June's 1.6-trillion-parameter LongCat-2.0, about four months later. LongCat-2.5 stands out for handling both image understanding and long development work in a single model, and for its low price of $1.20 per million output tokens.
On the other hand, the performance figures come from Meituan's own announcement, and the pricing is a limited-time discount. Since it is easy to try through OpenAI- or Anthropic-format APIs, it is a good idea to verify quality and cost on real work before adopting it. If you plan to use it for business, check its data handling and terms of use in advance as well.