Meituan launched a new AI model Monday1. The company calls it LongCat-2.5, built for coding, browsing and other jobs that run many steps in a row2.
A long-horizon job is one task that spans many smaller steps, like fixing a bug across ten files. The price beats Claude's by a wide margin on that kind of job. Meituan's own score against Claude still waits on an outside check, and the open weights Meituan promised months ago still sit unreleased.
Built to Dodge the Token Factory
Meituan made its fortune delivering food and running local errands, years before it touched frontier AI. Wang Xing runs the company as chief executive, and he keeps repeating one line at earnings calls.
"We will not become a Token factory. Our models and AI products will be used to support our core business, improving user and merchant experience as well as the company's operational efficiency," Wang Xing said in August3.
He framed the same point more broadly a moment later. "The long-term value of AI capabilities lies in whether they can truly integrate into real-world workflows, solve genuine problems, and develop into enduring, stable infrastructure," Wang Xing said3.
Money backs the words. Meituan's research spending grew 22.5 percent in the second quarter, to 7.7 billion yuan4. Wang Xing wants that spending to sort delivery routes and check a restaurant's margins, work most AI labs leave alone entirely.
Meituan's business focus still matters for a builder outside China. A model trained to watch warehouses and kitchens still writes code and reads screens the same way a chat assistant does. Meituan just built it for a different paying customer first.
The Price Undercuts Claude by a Wide Margin
LongCat-2.5 reads a context window of 1 million tokens2. A context window is how much text a model holds in memory at once, measured in chunks called tokens. This wide a window lets the model read an entire codebase in one sitting.
Meituan prices the model at 30 cents per million input tokens and $1.20 per million output tokens5. The company calls that a limited-time launch rate. Anthropic prices Claude Opus 5.5, its current flagship, at $4 per million input tokens and $20 per million output tokens8.
LongCat-2.5 charges 30 cents where Claude Opus 5.5 charges $4
Price per million tokens, Meituan's and Anthropic's own list prices, as of Nov. 3, 2026
Each pair of columns is one kind of token: the left column is LongCat-2.5's price, the right column is Claude Opus 5.5's price, for input tokens then output tokens.
LongCat-2.5's rate is a launch price Meituan calls limited-time, and it may rise once that window ends.
Sources: Startup Fortune [5]; Claude (Anthropic) [8]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | LongCat-2.5 | Claude Opus 5.5 |
|---|---|---|
| Input tokens | $0.30 | $4.00 |
| Output tokens | $1.20 | $20.00 |
A builder testing both models pays far less per token with LongCat. The gap holds even once the introductory rate ends. A team running thousands of agent calls a day would feel that difference on its monthly bill within a week.
Cheap access lowers the bar to try the model. Any developer can start today with a credit card and an API key alone.
Only Meituan Has Checked the Score
Meituan says LongCat-2.5 scores between Claude Opus 4.8 and Claude Opus 5 on its own mainstream tests2. The model also beats Opus 5 on some of them, the company says. Every one of those numbers comes from Meituan's own lab.
Meituan LongCat@Meituan_LongCatIntroducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active.
Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such… pic.twitter.com/KuZTk38HXr
BenchLM, a group that tracks published AI benchmarks, checked an earlier Meituan model against OpenAI's GPT-5.5 in August. It found zero benchmarks the two companies had tested in common. Artificial Analysis, an independent testing group, still carried zero LongCat-2.5 score as of Nov. 37.
Rui Ma reads the wider picture with more confidence. She founded Tech Buzz China, a research firm that tracks Chinese tech companies. "China is now training trillion parameter models on domestic hardware. Meituan's new 1.6T-parameter LongCat model was reportedly trained on around 50,000 Huawei Ascend chips," Ma wrote on X6.
Julian Lim reads this specific launch with more doubt. He covers AI models for Startup Fortune. "What is available is a cheap, capable API and a promise that the weights are coming. In a market this crowded, that promise is worth less than the file itself," Lim wrote5.
The Open Weights Keep Slipping
Open weights are a model's actual files. Anyone can download them and run them on their own machines. Meituan already knows how to post them. LongCat-2.0 launched June 30, and Meituan posted its full weights under an MIT license six days later10.
LongCat-2.0 opened its weights in 6 days; LongCat-2.5 still waits past 5 weeks
Days from each LongCat model's launch to its open-weight release, 2026
Each dot marks a dated event: gray dots are launch dates, and the single magenta dot marks the day LongCat-2.0's weights actually posted.
Sources: LongCat AI (Meituan) [2]; 36Kr [10]. Chart by Kelly Dennis.
The numbers behind this chart
| Date | Event |
|---|---|
| June 30, 2026 | LongCat-2.0 launches |
| July 6, 2026 | LongCat-2.0 weights post |
| Sept. 25, 2026 | 2.5-Preview launches. LongCat-2.5-Preview, API-only |
| Nov. 2, 2026 | 2.5 launches, API-only. LongCat-2.5's full release |
Any outside developer could run LongCat-2.0 on a private server that same week, keeping its data off the public internet entirely. LongCat-2.5-Preview arrived Sept. 25, with weights Meituan called coming soon2. The full LongCat-2.5 followed Nov. 2, still API-only, five weeks after that promise. LongCat-2.0 still holds the only weights the project has posted.
Rivals move faster on this point. Alibaba shipped Qwen3-Coder under an open Apache license the day it launched5. DeepSeek posted its own open coding agent, Super-Agent, the same week LongCat-2.5 arrived11.
What to do
Test LongCat-2.5 through its API this week. The price makes trying it almost free, and a sandbox test carries little risk.
Hold production work back from the model for the moment. Wait for Artificial Analysis or a similar outside lab to publish its own score first.
Ask about data retention before sending real customer data through the API. Meituan's own platform page still skips a stated retention policy and a service agreement7. A team that handles regulated data should treat that gap as a real blocker, past any benchmark score.
By the numbers
- 1.6 trillion: LongCat-2.5's total parameter count, with about 48 billion active per token2.
- $0.30: Meituan's launch price per million input tokens, against Anthropic's $4 for Claude Opus 5.558.
- 50,000: the Huawei Ascend chips Rui Ma says Meituan used to train the LongCat line6.
- 0 of 38: shared third-party benchmark scores BenchLM found between GPT-5.5 and an earlier LongCat model, as of August7.
- Six days: how long LongCat-2.0 took to go from launch to open weights, in July10.
- 22.5 percent: the year-over-year rise in Meituan's second-quarter research spending, to 7.7 billion yuan4.
What to watch
Track Artificial Analysis and similar labs for their first independent LongCat-2.5 score. Check GitHub and Hugging Face for the open weights Meituan promised in September. Either one would settle whether this launch holds up past Meituan's own tests.
Sources
- "Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks," TechNode, Nov. 2, 2026, https://technode.com/2026/11/02/meituan-launches-longcat-2-5-multimodal-ai-model-for-extended-coding-tasks/
- "LongCat-2.5 Native Multimodal Agent Model," LongCat AI (Meituan), Nov. 2, 2026, https://www.longcatai.org/
- "Meituan Does Not Operate a 'Token Factory' - Official Statement & Clarification," 36Kr, Aug. 29, 2026, https://eu.36kr.com/en/p/3964590925356547
- "Meituan Unveils New AI Model Trained on Homegrown Chips," Caixin Global, Nov. 3, 2026, https://www.caixinglobal.com/2026-11-03/meituan-unveils-new-ai-model-trained-on-homegrown-chips-102500305.html
- Julian Lim, "Meituan's LongCat-2.5 Jumps Into China's Open AI Coding Race," Startup Fortune, Nov. 2, 2026, https://startupfortune.com/meituans-longcat-25-jumps-into-chinas-open-ai-coding-race/
- Rui Ma, post on X, July 2, 2026, https://x.com/ruima/status/2072824914370048174
- "Meituan Releases LongCat 2.5 on Chinese Chips; Benchmark Claims Await Independent Tests," Tech Times, Nov. 3, 2026, https://www.techtimes.com/articles/329925/20261103/meituan-releases-longcat-25-chinese-chips-benchmark-claims-await-independent-tests.htm
- "Pricing," Claude (Anthropic), accessed Nov. 3, 2026, https://claude.com/pricing
- Meituan LongCat, post on X, Nov. 2, 2026, https://x.com/Meituan_LongCat/status/2117264311714537633
- "China's First Self-Developed Trillion-Parameter Large Model for Card Training and Inference Officially Open Source," 36Kr, July 6, 2026, https://eu.36kr.com/en/p/3884214101324041
- "DeepSeek open-sources Super-Agent, a coding agent that watches team chats for work," Help Net Security, Nov. 2, 2026, https://www.helpnetsecurity.com/2026/11/02/deepseek-dsh-superagent/
