Quick Answer
LongCat 2.5, released by Meituan on 2 November 2026, is a 1.6 trillion parameter native multimodal model built for long horizon agent tasks, priced on its official platform at a limited time discount of $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. A small coding service can be priced around those rates, but the discount has no published end date, the weights are not yet released, and no income is guaranteed.
Key takeaways:
- LongCat 2.5 was released on 1 to 2 November 2026, with general API availability on 2 November, and it is aimed at long horizon agent and coding tasks.
- It has 1.6 trillion total parameters, with about 48 billion active per token, a 1 million token context window, and up to 128,000 output tokens.
- The limited time discounted price is $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens, with no end date published.
- GitHub and Hugging Face releases were marked coming soon at launch, so today this is an API and chat model, not a self hosted model.
- A $149 package sold twice in a month is $298 in revenue in the illustration below, before time, taxes, fees, and token costs.
- This is a service business, not passive income. Clients, revisions, and support are the real work, and token costs are only one line in the maths.
What is LongCat 2.5, and why is it cheap per token?
LongCat 2.5 is Meituan's latest LongCat model, and its headline size is deliberately misleading if read alone. It has 1.6 trillion total parameters, but only about 48 billion are active for each token, because it is a mixture of experts model that routes each token to a subset of experts. That design is why a very large model can still be cheap to run: you pay for the active compute, not for waking the whole model on every token. The model is natively multimodal, which means it processes visual and text inputs in one token framework, and it is built for long horizon agent tasks, such as coding work that runs across many steps, visual coding, creative design, and 3D tasks. It was trained entirely on domestic Chinese accelerators, using W4A4 quantisation to handle 1 million token context workloads under 80 GB of memory. Those engineering details matter for developers, because they explain how Meituan can offer a 1 million token context at this price, but they do not change the seller's maths: you still pay per token, and you still need to manage context carefully. Availability at launch is through Meituan's online chat service and API platform, with Vercel AI Gateway also listing the model at the same $0.30 and $1.20 rates, which gives teams outside China a familiar route to try it. GitHub and Hugging Face releases were marked coming soon on 2 November, and no source checked confirms open weights for 2.5 yet. LongCat 2.0 remains the newest member known to have open weights. If your service depends on self hosting, LongCat 2.5 is not ready for that today. If your service runs through an API, it is ready to test now.
What does the price actually buy, token by token?
The official platform price is a limited time discount, and the platform itself says prices may change, with its own billing as the final word. That warning should be in every pricing model built on this rate. The table below uses the discounted rates as listed, and it shows why cached input matters so much for long agent sessions. | Per 1 million tokens | Price | What it means in practice | | Uncached input | $0.30 | First time the model reads a token, or a token that is not in cache | | Cached input | $0.006 | A token the model has already seen and cached, 50 times cheaper than uncached input | | Output | $1.20 | Every token the model writes, the expensive side of the bill | The cache gap is the most important number for an agent service. An agent that re-reads a whole repository on every step pays uncached input repeatedly. An agent that keeps context cached pays $0.006 per million for the repeated part instead of $0.30. Over a long task, that difference can be larger than the output bill. Design your service to reuse context where the platform allows it, and to avoid re-feeding the same large files on every step when a smaller, targeted input will do. A worked example helps. Suppose a small coding job uses 2 million uncached input tokens, 8 million cached input tokens, and 1 million output tokens. The token bill is 2 times $0.30, plus 8 times $0.006, plus 1 times $1.20, which is $0.60 plus $0.048 plus $1.20, or $1.848 in total. That is under $2 in tokens for a job that may take you hours of review, testing, and client communication. The token cost is not the constraint. Your time is. Now suppose the same job is run badly, with no caching, so all 10 million input tokens are uncached. The bill becomes 10 times $0.30 plus $1.20, or $4.20. Still small in absolute terms, but more than twice the well cached version, and the gap grows with every extra step. The lesson is not that LongCat 2.5 is free. It is that token discipline, not model choice, decides whether a cheap model stays cheap at scale. One more caution from the launch coverage: the Claude Code setup guide for LongCat 2.5, as written, compacts conversations at 200,000 tokens, well below the 1 million token context window. That is a setup default, not a model limit, but it shows why you should test the actual context behaviour of your tooling before promising a client that a 1 million token job will run untouched. A large context window only helps if your agent setup actually uses it.
What service should you actually sell?
Sell a narrow, repeatable service, not a general offer to do anything with AI. Three examples fit LongCat 2.5 well, because they use long context, repeated code reading, and reviewable output. First, a small codebase cleanup and documentation package. You take one repository, or one feature inside it, read it thoroughly, fix small bugs, remove dead code, and hand back a short report plus documentation on how it works. The client gets a codebase they can maintain, not a lecture about parameters. Second, a visual coding package for a small business site or internal tool. LongCat 2.5's native visual understanding fits work where the input includes screenshots, designs, or diagrams, such as turning a mockup into a working page, or fixing a layout bug that only shows up visually. Keep the scope to one page or one feature, with a clear before and after. Third, a prototype package for a small app feature. You build a working prototype, not a production system, with tests, a short handover note, and a clear list of what still needs hardening before launch. Prototypes are honest work, they are easier to price, and they let a client test an idea before committing to a full build. What these three share is a clear boundary. The client knows what they get, you know how long it should take, and the token bill stays inside a range you can predict. A vague offer, such as unlimited AI coding for a month, breaks all three, and it will break your pricing too, because a long horizon agent without a boundary can run for a very long time on a very large context.
What does the honest maths look like?
Use the illustration below as a way to test your own numbers, not as a promise of income. Every figure is labelled, and you should replace each one with your own costs before you quote a client. The setup is a seller offering a codebase cleanup package at $149 per package. In the illustration, the seller sells two packages in a month, works about 7 hours per package including client calls, reading code, fixes, testing, revisions, and handover, and keeps token spend inside the example above, at about $2 per package. | Line | Illustration | Notes | | Revenue, 2 packages at $149 | $298 | Illustration only, not a promise, clients are not guaranteed | | Token spend, 2 packages at about $2 | About $4 | On discounted LongCat 2.5 rates, well cached example above | | Other direct costs, such as tooling and payment fees | Not included in this illustration | Add your real fees, taxes, and tool subscriptions before quoting | | Time, 2 packages at 7 hours | 14 hours | Includes calls, reading, fixes, testing, revisions, and handover | | Revenue minus token spend, before time, tax, and fees | About $294 | $298 minus about $4, before tax, fees, and your time value | | Revenue per hour, before tax and fees | About $21.29 per hour | $298 divided by 14 hours, before tax, fees, and extra costs | Read that table carefully. The token bill is almost irrelevant at this price, which is the point of using a cheap model. What matters is whether you can find two clients, deliver two good packages in 14 hours, and get paid. If you cannot find clients, the token price does not save you. If you price too low, a cheap model does not save you either, because your time is still the main cost. Now test the other direction. If you raise the package to $249 and sell two in a month, revenue is $498, and revenue per hour before tax and fees is about $35.57 on the same 14 hours. If you sell only one package at $149, revenue is $149, and revenue per hour is about $10.64. The model price barely moves those numbers. Client volume and package price do. That is why this article spends more words on scope and selling than on tokens: tokens are cheap, clients are not. Also test the discount risk. The $0.30 rate is a limited time discount with no published end date, and the platform says prices may change. If the rate doubled tomorrow, the example token bill would rise from about $2 to about $4 per package, which would not break the illustration. If the rate rose tenfold, it would still be about $20 per package. A service priced at $149 can absorb a token price change far more easily than a service priced at $19. Price for your time, not for the current discount, and treat the discount as a margin bonus, not as the business model.
How do you avoid the traps?
The first trap is promising the 1 million token context as if it were a guarantee for every job. Test your setup first, check the compaction behaviour of your agent tool, and keep jobs inside a context size you have actually run. A job that compacts at 200,000 tokens in your setup is a 200,000 token job in practice, whatever the model card says. The second trap is assuming the weights are available. They are not, as of the launch coverage checked, and a service that depends on self hosting LongCat 2.5 today has no weights to host. Use the API or chat route, or pick a different model that actually ships weights, such as LongCat 2.0 or another open model, if self hosting is a hard requirement. The third trap is ignoring data and account rules. LongCat is a Meituan service, and teams outside China will often access it through a gateway such as Vercel AI Gateway, which adds its own terms, billing, and data handling. Read the gateway terms as well as Meituan's, check where data is processed, and do not feed client secrets, private codebases, or personal data into a service whose handling you have not checked. For sensitive client code, a local or private model may be the right choice even if it costs more per token. The fourth trap is overpromising on quality. LongCat 2.5 is a new model with 25 official benchmarks reported in its launch material, but benchmarks are not client results. Test it on your own typical jobs, compare it with a model you already trust on the same job, and only sell work you have actually delivered and reviewed. A cheap model that produces code you do not check is not a bargain. It is a liability. For readers building a freelance offer around AI tools, our research agent service guide on DeepSeek Search and our agent service guide on Nebius credits use the same honest maths on different rates, and our open source LLM pricing guide helps compare LongCat with models you can self host today. Sources: Meituan LongCat 2.5 announcement, 1 November 2026, and LongCat API platform changelog and pricing page, 2 November 2026, as reported by TechNode, 2 November 2026, The Dev Drift, and APIYI, 2 to 3 November 2026; Vercel AI Gateway listing at the same $0.30 and $1.20 rates, as reported 2 November 2026; GitHub and Hugging Face status marked coming soon, as reported 2 November 2026.
Frequently Asked Questions
What is LongCat 2.5?
It is Meituan's native multimodal model for long horizon agent tasks, released on 1 to 2 November 2026, with 1.6 trillion total parameters, about 48 billion active per token, a 1 million token context window, and up to 128,000 output tokens.
How much does LongCat 2.5 cost?
On its official platform, at a limited time discount, $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. No end date for the discount was published in the coverage checked, and the platform says prices may change.
Can I download and self host LongCat 2.5?
Not as of launch. GitHub and Hugging Face releases were marked coming soon on 2 November 2026, and no source checked confirms open weights for 2.5. Use the API or chat route for now, or pick a model that actually ships weights if self hosting is required.
Is LongCat 2.5 really cheaper than other coding models?
Per token, the discounted rate is low, and cached input is 50 times cheaper than uncached input. Whether it is cheaper for your job depends on your cache use, your output volume, and whether your agent setup actually uses the full context window.
What service should I sell with it?
A narrow, repeatable package, such as a codebase cleanup, a visual coding fix for one page, or a prototype for one feature, priced by package with scope, revisions, and delivery format stated upfront.
Will selling two $149 packages a month make me a living?
No, on the illustration in this article, two packages give $298 in revenue and about $21.29 per hour before tax and fees, on 14 hours of work. Treat that as a starting test, not an income promise, and raise price or volume only after real clients confirm the offer.



