Meituan makes LongCat-2.5 generally available: a low-cost multimodal model for coding agents
Meituan has released LongCat-2.5 on its API platform, a multimodal model built for long coding and agent tasks at $0.30 per million input tokens. What it does, what it costs and what to check before testing it.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
Meituan, the Chinese delivery and local-services company, made its LongCat-2.5 model generally available on its developer platform on 2 November 2026, according to the LongCat API platform changelog. It follows a preview released on 25 September. Meituan is pitching the model at long, multi-step coding and agent work, and its list price sits well below most frontier models from US labs.
What it actually does
The changelog entry for 2 November describes three things:
- Native multimodality. The model takes images as well as text, and Meituan says it combines visual understanding with coding for visual coding, creative design and 3D tasks.
- Long-horizon agent work. It is meant to handle large codebases and whole-project builds through multi-step reasoning, tool use and end-to-end task execution. "Long-horizon" simply means a task that takes many steps and a long time, rather than a single question and answer.
- Works with existing agent tools. Meituan says it is compatible with Claude Code, Codex, Hermes and other agent harnesses. A harness is the program that runs the agent loop, reads files and calls tools; the model is the part that does the thinking.
When Meituan launched the preview, its announcement on X described it as a 1.6-trillion-parameter model with about 48 billion parameters active at a time, a 1-million-token context window, and native multimodality, aimed at work across terminals, browsers, GUIs, spreadsheets and design tools. A 1-million-token context window means the model can take in very large inputs at once, such as a big chunk of a codebase or a long set of documents. Meituan has not published a separate spec sheet for the general release in the pages we read, so treat those figures as the preview's.
What it costs
The LongCat pricing page lists LongCat-2.5 as the only model on its pay-as-you-go plan, at these limited-time discounted prices per million tokens:
| Item | USD | CNY |
|---|---|---|
| Input (not cached) | $0.30 | ¥2 |
| Input (cached) | $0.006 | ¥0.04 |
| Output | $1.20 | ¥8 |
Tokens are the small chunks of text a model reads and writes; a million tokens is hundreds of thousands of words. "Cached" input is text the model has already seen in a recent request, such as the same codebase sent again, which is billed at a much lower rate. The page gives no end date for the discount and notes that prices may change, with the platform's own billing being the final word.
The preview is also listed on Vercel's AI Gateway at the same $0.30 and $1.20 rates, with a context window of about 1 million tokens and up to 131,072 output tokens, which gives teams outside China a way to try it through a familiar provider.
Who should care
This matters most to teams already running coding agents that pay by the token. If your developers use an agent harness that lets you swap the model behind it, a cheap model with a large context window is worth testing on routine work such as refactors, test writing and documentation, where cost adds up quickly.
For business owners, the wider point is that competition on price for agent-grade models keeps coming from outside the usual US labs. That puts pressure on what you pay for AI tooling, even if you never use this model directly.
Limits and caveats
- No independent benchmarks yet. The changelog and pricing pages make no benchmark claims. Judge it on your own tasks, not on headlines.
- Data handling. The service is run by a Chinese company. Before sending client code or personal data, check the platform's terms, where data is processed and how long it is kept, and whether that fits your contracts and GDPR or sector rules.
- Prices are promotional. The listed rates are marked limited-time, with no end date.
- Weights. The pages we read describe API and chat access only; we found no download of the model weights for LongCat-2.5, so you cannot self-host it today.
What to do this week
- If you run coding agents, pick two or three routine tasks from your backlog and run them on LongCat-2.5 alongside your current model in a test repository with no sensitive code. Compare results, review time and token cost.
- Ask your legal or security lead whether a China-hosted model API is acceptable for any of your data, and write the answer down so the question doesn't come up again with every new model.
- If you route models through a gateway, check whether it already lists LongCat, so testing doesn't require a new vendor account.
The DevDrift helps teams put AI coding agents into their workflows with review steps and cost tracking, whichever model sits behind them.
