# What is LongCat-2.5: Meituan's open 1.6T agent model, and its price

> The MIT-licensed model reads 1 million tokens, was trained on Chinese chips, and costs $0.30 per million input tokens on Meituan's API for now.

URL: https://claudegpt.si/news/what-is-longcat-2-5  
Published: 2026-11-15T10:57:38.415Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Models  
Topics: Meituan, LongCat, Open-weight models, Language models, AI chips  
Publisher: ClaudeGPT Si

## Key takeaways

- LongCat-2.5 is a 1.6-trillion-parameter mixture-of-experts model from Meituan that uses about 48 billion parameters per token.
- Its weights are on Hugging Face under the MIT license, and Meituan says it was trained entirely on Chinese-made chips.
- Meituan's API charges a limited-time $0.30 per million input tokens and $1.20 per million output tokens, as of November 15, 2026.

LongCat-2.5 is a large AI model from Meituan, the Chinese food delivery and local services company, built to work as an agent: writing and fixing code across whole projects, browsing the web and using tools over many steps. It has 1.6 trillion parameters, reads up to 1 million tokens, and its weights are on [Hugging Face](https://huggingface.co/meituan-longcat/LongCat-2.5) under the MIT license, so anyone can download, run and change it, including for commercial use.

Two things make LongCat-2.5 worth knowing about. Meituan's own benchmark table puts it close to the best closed models from OpenAI and Anthropic on several agent tests. And Meituan says it was trained entirely on Chinese-made accelerators, not Nvidia GPUs. The model's own site says, in Chinese, that its training was done entirely on domestic chips.

## What LongCat-2.5 is

The [model card](https://huggingface.co/meituan-longcat/LongCat-2.5) calls it a "native multimodal" mixture-of-experts model. Mixture of experts means the model is split into many smaller sub-networks and only some of them run for each word. LongCat-2.5 has 1.6 trillion parameters in total but uses about 48 billion for each token, which keeps it much cheaper to run than its size suggests. (Our explainer on [mixture of experts](https://claudegpt.si/news/what-is-mixture-of-experts) covers how that works.)

It takes text, images and video as input and answers in text. Images are turned into tokens that sit in the same space as words, an approach Meituan carried over from its earlier LongCat-Next research. The model has two modes: a thinking mode, on by default, that reasons before answering, and a faster non-thinking mode that uses fewer tokens.

Meituan says it built the model for "long-horizon agentic tasks": working in large codebases, building whole projects, design work and 3D tasks. It follows LongCat-2.0 and an earlier LongCat-2.5-Preview, which Meituan put on its API on September 25, 2026.

## When it came out

Meituan's [API change log](https://longcat.chat/platform/docs/change-log) lists "LongCat-2.5 Now Available" under version 2026-11-02, so developers have been able to call it since November 2. The Hugging Face repository with the weights was created on November 4. The Decoder [reported on November 15](https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/) that the weights went out a week after the model was unveiled.

## How it scores

Meituan's model card compares LongCat-2.5 with GPT-5.5, GPT-5.6 Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash. Meituan ran the LongCat numbers itself and calls its in-house results "indicative only"; most competitor scores are cited from public sources.

| Benchmark (what it tests) | LongCat-2.5 | Claude Opus 5 | GPT-5.6 Sol | DeepSeek-V4.1-Flash |
| --- | --- | --- | --- | --- |
| Terminal Bench v2.1 (agent work in a terminal) | 88.4 | 89.1 | 88.8 | 90.6 |
| Terminal Bench v4.0 (harder version) | 38.5 | 53.9 | 37.3 | 31.2 |
| DeepSWE v1.1 (fixing real software issues) | 64.6 | 68.8 | 72.7 | 74.2 |
| BrowseComp (finding facts on the web) | 91.9 | 90.8 | 90.4 | 91.5 |
| Humanity's Last Exam, with tools | 63.1 | 64.7 | 58.0 | 63.9 |
| MCP Atlas (using tools over MCP) | 83.1 | 85.8 | 81.8 | 84.4 |

The pattern is consistent. LongCat-2.5 is near the top on web search and close on most agent tests, but Claude Opus 5 leads clearly on the hardest coding benchmarks, and DeepSeek's newest model beats it on the software engineering tests. The jump from LongCat-2.0 is large: on DeepSWE it went from 9.7 to 64.6, and on Terminal Bench v4.0 from 3.0 to 38.5.

No independent lab had scored LongCat-2.5 when this was written. Artificial Analysis had no page for it, and OpenRouter listed only LongCat-2.0. Until outside tests arrive, treat the table as Meituan's claims.

## Built without Nvidia

The model card says LongCat-2.5's training and serving run on "domestic AI accelerators." The deployment guide in Meituan's [SGLang-FluentLLM repository](https://github.com/meituan-longcat/SGLang-FluentLLM) targets NPUs, a class of AI chips, rather than Nvidia GPUs. According to The Decoder, the sample setup uses 194 Huawei Ascend accelerators.

Meituan also shipped a [4-bit version](https://huggingface.co/meituan-longcat/LongCat-2.5-W4A4) that stores both the experts' weights and their activations in 4 bits. Against an 8-bit setup, Meituan reports 43.3% higher throughput when reading prompts and 42.77% higher when generating, with accuracy it describes as on par.

This matters beyond Meituan. US export rules limit which Nvidia chips Chinese companies can buy. A frontier-class open model trained on Chinese hardware is evidence that those limits slow Chinese labs less than intended, though Meituan has not said how many chips or how long training took.

## How much LongCat-2.5 costs

The weights are free to download. Running a 1.6-trillion-parameter model yourself takes a large cluster, so most people will use Meituan's API instead. As of November 15, 2026, the [LongCat-2.5 pricing page](https://longcat.chat/platform/docs/pricing/longcat-2.5) lists these limited-time prices per million tokens:

| | Price per 1M tokens |
| --- | --- |
| Input | $0.30 |
| Cached input | $0.006 |
| Output | $1.20 |

The page gives no end date for the offer and no regular price. For comparison, Google's new [Gemini 4 Hydrogen](https://claudegpt.si/news/gemini-4-hydrogen-fast-model-pricing) costs $1.50 per million input tokens.

## How to use it

The [API docs](https://longcat.chat/platform/docs/) list a 1-million-token context window and up to 256,000 tokens of output. The API accepts both the OpenAI format (Chat Completions) and the Anthropic format (Messages), so most existing code and tools work by changing the base URL and the key. Meituan publishes setup guides for Claude Code, Codex, OpenCode, Cline, Kilo Code, Hermes Agent and others. There is also a chat app at [longcat.ai](https://longcat.ai/).

To run the open weights, the model card points to SGLang-FluentLLM on NPU hardware. No hosting provider on Hugging Face served the model yet.

## What to watch

The next step is outside testing. Independent scores from Artificial Analysis or SWE-bench will show whether Meituan's numbers hold, and the regular API price will show whether $0.30 per million input tokens lasts beyond the launch offer.

## Frequently asked questions

### Is LongCat-2.5 open source?

Its weights are published on Hugging Face under the MIT license, which allows commercial use, changes and redistribution. The license grants no rights to Meituan's trademarks or patents.

### How much does the LongCat-2.5 API cost?

As of November 15, 2026, Meituan lists limited-time prices of $0.30 per million input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens. It gives no end date or regular price.

### Can I use LongCat-2.5 with Claude Code?

Yes. The API accepts both OpenAI and Anthropic request formats, and Meituan publishes setup guides for Claude Code, Codex, OpenCode, Cline and other tools.

### Who makes LongCat?

LongCat is built by Meituan, the Chinese food delivery and local services company. Its earlier models include LongCat-Flash and LongCat-2.0.

## Sources

- [LongCat-2.5 model card](https://huggingface.co/meituan-longcat/LongCat-2.5), Meituan
- [LongCat API platform quick start](https://longcat.chat/platform/docs/), Meituan
- [LongCat-2.5 pricing](https://longcat.chat/platform/docs/pricing/longcat-2.5), Meituan
- [LongCat API change log](https://longcat.chat/platform/docs/change-log), Meituan
- [LongCat-2.5-W4A4](https://huggingface.co/meituan-longcat/LongCat-2.5-W4A4), Meituan
- [Introducing LongCat-2.5](https://longcat.ai/blog/longcat-2.5/), Meituan
- [Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU](https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/), The Decoder
