LongCat-2.5: 1.6T Parameters, 1M Context, $0.30 API

By Muzamil Ahmad · Updated 2026-11-03 · Coding

#coding#longcat 2.5#meituan longcat#longcat api pricing#ai coding agents#1m context model#chinese ai models

Save

How we write our guides: we start from the vendor’s own announcement or documentation, not from other “best of” lists. Launch dates and features are checked against that primary source; anything reported only by coverage is labelled as reported. Prices and allowances are given only when the vendor publishes them — otherwise we say “not published” rather than guessing. We do not invent test results, prices, or rankings. Our guides are drafted with AI assistance, then reviewed and edited before publication; the final call on what we recommend is a human one.

# LongCat-2.5: Meituan's 1.6-Trillion-Parameter Model Is Built to Finish Projects, Not Chats

The most interesting AI lab of 2026 might be a food-delivery company. Meituan — the Chinese platform that delivers more meals than almost anyone on Earth — has spent two years quietly building a frontier model programme, and on November 2, 2026 its LongCat team made LongCat-2.5 generally available: a 1.6-trillion-parameter, natively multimodal model with a 1-million-token context window, priced at $0.30 per million input tokens. That is frontier scale at budget-model prices, aimed squarely at the longest, messiest job in AI right now: autonomous agents that work inside real codebases for hours.

This guide covers what was actually released, the architecture change that matters (vision turned into tokens), Meituan's benchmark claims read honestly — including the rows where a rival wins on Meituan's own table — the pricing with its discount catch, how to try the model today inside tools like Claude Code, and the two things it is not yet: open weights, and independently tested.

What Meituan announced on November 2

LongCat-2.5 reached general availability on the LongCat API Platform on November 2, five weeks after a preview build landed on September 25. The positioning shifted between versions. LongCat-2.0, which Meituan open-sourced in June 2026 with MIT-licensed weights, was an agentic coding model. The 2.5 pitch is broader: a model for long-horizon agentic tasks — work that spans terminals, browsers, graphical interfaces, spreadsheets and design tools, and whose unit of success is not a good answer but a finished project.

Three changes headline the release notes. First, vision became native (the next section unpacks how). Second, the reinforcement learning got longer: Meituan describes training on synthetic environments mined from real GitHub issues and pull requests, with fail-to-pass verification, asynchronous rollouts, and mixed training across different agent harnesses — the model practised inside the same kinds of scaffolding it will be used in. Third, the infrastructure absorbed the cost: the model was trained and is served on domestic Chinese accelerators, with W4A4 inference (4-bit weights and 4-bit activations) and engineering work on keeping training and inference numerically consistent, which matters when a model is refined by its own rollouts.

Access, as of publication: the model is live on Meituan's API platform and its longcat.ai chat service, and it is listed on gateways including Vercel AI Gateway and AnyRouter (bring your own LongCat key). It is not yet on OpenRouter, whose Meituan listing still shows LongCat-2.0 only, and — the catch this guide returns to — the GitHub and Hugging Face links on the launch pages are marked "coming soon."

The numbers that matter

Two of those numbers deserve a pause. The 48-billion active count is how a 1.6-trillion-parameter model can be served at $0.30 a million: only a small slice of the network fires for any given token, so the compute bill resembles a mid-size model's while the knowledge capacity resembles a giant's. And the 1M context is not a stretched version of a smaller window — the LongCat line was built around sparse attention (LongCat Sparse Attention in the 2.0 generation) designed for million-token sequences, which is why Meituan prices a 900K-token request at the same per-token rate as a 9K one. There is no long-context surcharge tier.

For the competitive frame: Moonshot's Kimi K3.1 Pro offers the same 1M context at $2/$10 per million — roughly seven times LongCat's input price — while DeepSeek's V4.1 Flash matches LongCat's price almost exactly. LongCat-2.5 is not the only cheap agent model; it is the biggest one at the price.

Why a food-delivery company builds frontier models

The LongCat programme is older than this launch suggests. Meituan open-sourced LongCat-Flash in September 2025 — a 560-billion-parameter mixture-of-experts model released with its weights, at a time when most Chinese labs were still treating openness as a marketing phase. LongCat-2.0 followed in June 2026 at 1.6 trillion parameters with MIT-licensed weights, and the 2.5 preview arrived that September. Three flagship generations in fourteen months, from a company whose public business is delivering dinner.

The logic is not mysterious. Meituan operates one of the highest-volume consumer platforms on Earth — ordering, dispatch, customer service, merchant tooling — and every one of those workloads is now an AI workload. A lab built inside that demand has two advantages money cannot quickly buy: real agent traffic to learn from, and a parent whose compute budget is justified by operations rather than fundraising. The domestic-accelerator programme belongs to the same picture: training and serving a 1.6-trillion-parameter model on Chinese silicon, with 4-bit inference engineered in from the start, is a supply-chain strategy as much as a research one. When the release notes emphasize W4A4 and train–inference consistency, they are describing a model designed to be cheap to run at national scale, which is precisely what its API price then advertises to the rest of the world.

That timeline also frames the one regression in this launch: the programme that built its reputation on open weights has, for now, stopped releasing them. Whether 2.5's weights follow its predecessors' is the single most consequential unanswered question about this model.

Vision as tokens: the architecture change that matters

Most multimodal models bolt a vision encoder onto a language model and translate between the two. LongCat-2.5 instead converts visual signals into multi-level discrete tokens — images and video frames are quantized into the same kind of token IDs the language backbone already processes, at native resolution. One sequence type, one backbone, no separate visual pathway to fine-tune around.

Why should a user care about an internal representation choice? Because it changes what the model can be asked to do in one pass. A screenshot of a broken UI, a photo of a whiteboard, a short clip of a failing build — each becomes part of the same working memory as the code and the error log, rather than an attachment the model describes and then forgets the details of. Meituan's stated use cases follow from that design: visual coding (build the interface from the mockup), creative design work, and 3D generation, alongside the terminal-and-browser agent work. The release includes a text-to-3D benchmark, 3DCodeBench, that exists precisely because the vision pathway is meant to be productive, not just descriptive.

The honest limit: discrete visual tokens trade fidelity for unity. A model that reads images as tokens may miss fine visual detail that a dedicated high-resolution vision encoder would catch, and until independent testers publish results, "native multimodal" is a design claim with demos behind it, not a measured advantage. Our best AI coding assistants guide covers the tools this class of model plugs into once the agent work matters more than the architecture.

The benchmarks: Meituan's claims, read honestly

Every score in this section is Meituan-reported. As of November 3, Artificial Analysis lists only LongCat-2.0 and LongCat Flash Lite — no independent body has scored 2.5 yet. With that label firmly attached, the release-notes table (2.5 against its own predecessor) is still instructive, because the deltas show where the training went:

The pattern is consistent: the largest jumps sit exactly where Meituan said the reinforcement learning went — long, verified, multi-step engineering tasks (DeepSWE up 55 points, Terminal-Bench v4.0 up from near-zero) rather than chat quality. In Meituan's cross-model comparisons, LongCat-2.5 scores 91.9 on BrowseComp, which the company ranks first in its comparison set, and its AutomationBench result sits ahead of Claude Opus 5's 50.3 in the same table.

Now the part launch posts usually skip. Reviewing Meituan's own published table, the sharpest early analysis found that DeepSeek V4.1 Flash — which costs the same $0.30/$1.20 — wins most of the head-to-head coding rows (our DeepSeek V4.1 Flash guide covers that model). LongCat's clearest wins are in browsing and long-horizon agent suites, not raw code generation. Add the standard caveats — vendors choose harnesses, context lengths and comparison sets, and a benchmark run inside a vendor's scaffold flatters the home model — and the fair summary is: LongCat-2.5 looks genuinely strong at agent endurance, competitive at coding, and unproven everywhere until a third party runs it. That is still a remarkable position for a $0.30 model; it is just not the coronation the headline numbers suggest.

Pricing: the cheapest frontier-adjacent API right now

The pay-as-you-go prices, per million tokens, at the launch's limited-time discount:

Three details make this better than the headline. Cached input costs 2% of uncached — for a coding agent that resends the same repository context every turn, most of the input bill effectively drops fifty-fold. Only successful calls are billed (HTTP 200s), and prepaid balance does not expire. And there is no length tiering: a near-million-token request pays the same rates as a short one. Meituan's own worked example prices a heavy agent request — 600,000 uncached input tokens, 400,000 cached, 500,000 output — at $0.7824. Under a dollar for a request that would cost several dollars at Kimi's rates and tens of dollars at flagship Western prices (our Claude review lists what the Opus line charges).

The catch is printed in the same table: these are limited-time discounted rates with no published end date, and Meituan has not published the undiscounted list price for 2.5 alongside them. The older 2.0 pricing suggests list rates could be roughly two to two-and-a-half times higher when the discount ends. Budget for the discount while it lasts; do not build a business plan that only works at $0.30.

How to try it today

Through the API platform. Create a key on Meituan's LongCat API Platform, then call the model in either the OpenAI or the Anthropic request format — the dual compatibility is deliberate, and it means existing code written for either ecosystem usually needs only a base-URL and key change. The model ID on gateways is meituan/longcat-2.5.

Inside your agent harness. Meituan states compatibility with Claude Code, Codex, Hermes, OpenClaw, OpenCode and Kilo Code. In practice that means pointing the harness at the LongCat endpoint as the model provider — the cheapest meaningful experiment in AI right now is running your usual coding agent on your own repository at these prices for a day and comparing the result with your default model. Existing verified platform users were also granted free token balances around the launch windows, so an older LongCat account may already hold credit.

Through a gateway. Vercel AI Gateway carried the preview at the same $0.30/$1.20 rates — the familiar-provider route for teams outside China — and AnyRouter lists the GA model for bring-your-own-key use. OpenRouter does not list 2.5 yet; if your stack routes everything through it, the direct API or a gateway is the way in for now. The release notes and benchmark table and the most sceptical early review we found, Eesel's LongCat 2.5 explainer, are the two reads worth your time before you commit a workload.

Who should test it first

Agent-heavy development teams get the clearest proposition. If your coding agent's monthly bill is dominated by long sessions — repository-wide refactors, test-suite repairs, migration work — a same-format API at a seventh of Kimi's price with a 1M window is a one-afternoon experiment with a measurable answer: run your ten hardest recent tasks on both and compare outcomes per dollar, not per token.

Multimodal pipeline builders are the second natural audience. If your workflow currently chains a vision model into a language model — screenshot in, description out, code after — a single backbone that holds the image, the code and the logs in one context removes a lossy hand-off, and the visual-coding demos are aimed at exactly you.

Everyone else should wait for one of two signals: an independent scoreboard entry, if benchmark certainty is what your procurement needs; or the weights, if self-hosting is the point. Neither exists today, and a model this cheap to test will still be cheap to test when they arrive. What does not justify waiting is the brand: by November 2026, "a Chinese lab I had not heard of" describes half the models at the top of the public leaderboards, including several covered elsewhere on this site.

The catches, in full

No weights, and no date. LongCat-2.0 shipped MIT-licensed weights; 2.5's GitHub and Hugging Face links say "coming soon," and Meituan has announced no open-weight release or timeline for this version. If your interest in LongCat is self-hosting, today there is nothing to download — the contrast with the Mistral Large 4 release, whose weights arrived with a (restrictive) licence on day one, is the whole story of open AI in late 2026: "open" now describes a spectrum, and 2.5 currently sits at the API-only end of it.

No independent verification. Every benchmark above is the vendor's. The model may well be as strong as claimed — the 2.0 generation tested respectably once outsiders ran it — but as of November 3 there is no third-party scoreboard entry for 2.5, and buying decisions for production workloads should wait for one or run on your own evaluations.

The discount cliff. Covered in the pricing section: today's rates are promotional, the end date is unpublished, and the list price is unpublished. Treat the gap as a known unknown.

Platform practicalities. The API platform is a Chinese provider's platform: international payment support has historically lagged on LongCat's platform, documentation quality varies by page, and your prompts transit Meituan's infrastructure under its data policies. None of that is disqualifying — it is the same diligence any new provider deserves, applied to one whose consumer brand most Western developers have never heard of.

LongCat-2.5 vs the field

Against Kimi K3.1 Pro ($2/$10, 1M context, weights promised by November 4 — our Kimi weights guide tracks that drop): Kimi has independent scores placing it near the frontier and an open-weights commitment; LongCat is a seventh of the price with vendor-only scores and no weights date. Against DeepSeek V4.1 Flash (same $0.30/$1.20): DeepSeek wins more coding rows on Meituan's own table and has the longer public track record; LongCat counters with the bigger context story, native vision and the agent-endurance results. Against the Western flagships, LongCat is not really competing on peak quality at all — it is competing on the price of good enough, for hours at a time, which is where agent workloads actually burn money.

Put concrete arithmetic on the gap, using headline rates with no caching benefit on either side. A heavy agent day — 50 million input tokens, 5 million output — costs about $21 on LongCat-2.5 (50 × $0.30 + 5 × $1.20). The same day at Kimi K3.1 Pro's $2/$10 costs about $150. At Claude Opus-class prices it costs several hundred. Caching narrows or widens those gaps depending on the workload, and DeepSeek matches LongCat's line exactly — but the order of magnitude is the point: a week of autonomous development on LongCat costs roughly what a single day costs on the model it is benchmarked against. Prices at that ratio do not compete with the frontier; they change what the frontier has to justify.

The strategic picture is the one this site keeps returning to: Chinese labs are shipping frontier-scale models at prices that reset what an agent hour should cost, and the pressure lands on everyone else's pricing page. Whether LongCat-2.5 specifically becomes your default matters less than what its existence does to the market you buy from.

Verdict

LongCat-2.5 is the price-performance story of November so far: frontier scale, a genuine architecture idea in vision-as-tokens, agent training aimed at finishing work rather than chatting about it, and API pricing that makes a day of autonomous coding cost less than lunch. It launches with the two honest asterisks of the season — vendor-only benchmarks and weights that exist as a "coming soon" link — and one financial asterisk, a discount with no visible expiry. Test it now while the testing is cheap; commit to it when the independent numbers land.