Tools / Analysis
LongCat 2.5 puts images, video and a million-token workspace inside a coding agent
Meituan’s new model is available through APIs. The useful combination is visual input, a large working context and compatibility with existing coding tools—not the trillion-parameter headline alone.
LongCat 2.5 can take a screenshot or a video alongside a written request, then work through a coding task with a million-token context. Meituan introduced the model in a post dated 1 November; Vercel added it to AI Gateway on 2 November. For developers, that creates another model to try inside an existing agent workflow, with visual material included in the brief.
1.6 trillion parameters; about 48 billion active
Meituan describes a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token. In this architecture, routing selects parts of the network for each token instead of using every expert every time. Total capacity and active computation are different quantities. Neither number, on its own, establishes answer quality or the hardware needed to serve the entire model.
Sources: Meituan LongCat [1]
The brief can include what you see
The native API documents both image and video input. Images can arrive as encoded files or links; video is supplied by URL. This is visual understanding, not a claim that LongCat generates video. Meituan lists tasks such as reading screenshots and interpreting charts. Vercel separately emphasizes coding, tool use and full-project work. Together, those capabilities make a concrete workflow possible: give an agent the visual reference as well as the code it needs to change.
That still leaves an important distinction: accepting a reference does not prove the model will reproduce its layout accurately. A useful evaluation would inspect the resulting interface, its behavior and its code—not just the model’s description of the picture.
Sources: LongCat API Docs [2] · Vercel [6]
A larger workspace, not permanent memory
The parameter reference specifies 1,048,576 tokens of context, counting input and output together. That is space available within a request; it does not establish perfect recall across the whole window or persistent memory between sessions. For a coding agent, the practical question is whether it can keep using the relevant files and previous decisions as the task grows.
Meituan exposes OpenAI-compatible and Anthropic-compatible API formats under the model name LongCat-2.5. Vercel uses meituan/longcat-2.5 and documents setup for coding clients including Claude Code, Codex and Cursor. These are integration routes, not evidence of equal behavior across different agent harnesses.
Sources: LongCat API Docs [4] · LongCat API Docs [3] · Vercel [6]
The price is an invitation to compare
The official limited-time rates are $0.30 per million uncached input tokens, $0.006 for cached input and $1.20 for output. For long agent runs, token price is only part of the bill: repeated attempts, unnecessary tool calls and human corrections also matter. The useful comparison is the cost of a completed, checked task. LongCat 2.5 now gives developers an accessible API route to make that comparison on their own work.
Sources: LongCat API Docs [5]
SOURCES & METHOD
The record behind the article
Checked against Meituan’s published model description, API documentation and pricing, and Vercel’s 2 November availability notice. Capabilities are vendor claims; we have not run the model or reproduced its evaluations. The API documentation gives conflicting maximum-output limits, so this article does not quote one.
- LongCat-2.5: Native Multimodality for Agentic Tasks Meituan LongCat · 2026-11-01
- Visual Understanding LongCat API Docs
- LongCat API Platform Quick Start Guide LongCat API Docs
- Model Parameters LongCat API Docs
- LongCat-2.5 API pricing LongCat API Docs
- LongCat 2.5 now available on AI Gateway Vercel · 2026-11-02
