LongCat-2.5 review: price, open weights and benchmarks

This review is built from what could be checked on November 3, 2026: Meituan's LongCat-2.5 technical blog, API documentation, pricing pages, change log, terms and privacy policies; Meituan's Hugging Face and GitHub pages; Meituan's news release and coverage by Xinhua, TechNode and IT之家; commentary by Startup Fortune and the Chinese site 什么值得买; ZenMux's and Artificial Analysis's listings; and the price pages of five rival models. LongCat-2.5 is two days old and no outside lab had published a benchmark run of it, so every score below is Meituan's own unless another source is named.

What LongCat-2.5 is

LongCat-2.5 is a large language model from LongCat, the AI team of Meituan, the Beijing-based platform for food delivery and local services. The LongCat site, chat and API are run by Beijing Kuxun Interactive Technology Co., Ltd. and its affiliates (API terms). Meituan describes LongCat-2.5 as a native multimodal model for long-horizon agent tasks: a mixture-of-experts model with 1.6 trillion parameters, about 48 billion of them active for each token, built on LongCat-2.0 (technical blog).

"Native multimodal" here means the model turns images into discrete tokens and reads them in the same sequence as text, instead of passing them through a separate vision model. The method follows Meituan's earlier open model LongCat-Next. The output is text: code, answers and reports. LongCat's documentation lists no audio input, and the model does not draw images.

Next to its predecessor, from Meituan's two technical blogs and the LongCat-2.0 model card:

Sources: LongCat-2.5 blog, LongCat-2.0 blog, LongCat-2.0 model card, model parameters.

What happened on November 2

Meituan released LongCat-2.5 on November 2, 2026, five weeks after LongCat-2.5-Preview reached its API on September 25 with image understanding (change log, TechNode on the Preview). The technical blog is dated November 1. Xinhua, TechNode and IT之家 all reported the launch with the same core figures (Xinhua, TechNode, IT之家).

Meituan's own release makes three claims: that LongCat-2.5 was trained entirely on Chinese-made computing hardware, that on many benchmarks its scores fall between Claude Opus 4.8 and Claude Opus 5 and on some beat Opus 5, and that anyone can try it on longcat.ai while developers use the API launched the same day (Meituan news release). Xinhua quoted Meituan's chief executive Wang Xing from the August 28 earnings call saying the company would not become a "token factory" and that its models serve its main business; Meituan's merchant workbench Miaoshou runs on LongCat models.

What did not ship: the weights. The blog marks GitHub and Hugging Face as coming soon, and on November 3 neither Meituan's Hugging Face nor its GitHub organisation had a LongCat-2.5 entry. That is a change from LongCat-2.0, which Meituan announced as open source in its launch post. Startup Fortune's verdict is that until the files are posted, the openness belongs to the older model (Startup Fortune).

How you use it

LongCat-2.5 is the only model in the API quick start on November 3. The API speaks three formats at api.longcat.ai: OpenAI Chat Completions and Responses, and Anthropic Messages, all with the model name LongCat-2.5 (quick start). That makes it a drop-in backend for agent tools; LongCat's documentation has setup pages for Claude Code, Codex, OpenCode, Cline, Kilo Code, Hermes Agent, Tencent's CodeBuddy and WorkBuddy, DeepSeek's own harness and others. For Claude Code the change is a few lines in settings.json (Claude Code guide).

Limits per request, from the model parameters and visual understanding pages:

• context window of 1,048,576 tokens, input and output together, and up to 262,144 output tokens;
• up to 100 images in PNG, JPG, JPEG, WebP or GIF, 10 MB each, as Base64 or a link;
• up to 10 videos by public URL only, each up to 100 MB and under 3,600 seconds;
• thinking switched on or off per request, with its own budget of up to 262,144 tokens;
• temperature fixed at 1.0 with thinking and 0.7 without; other values are overridden, not rejected.

The last point matters for anyone who relies on low temperature for repeatable code or extraction: you cannot set it.

What it costs

LongCat sells LongCat-2.5 by the token, from a prepaid balance. Prices on the pricing page on November 3, 2026, in USD per million tokens:

Both the English and the Chinese pricing page call these limited-time prices for the platform's launch, and neither gives a regular price. Two outside listings hint at it: ZenMux, which resells LongCat-2.5-Preview, shows that model's price before a 60% discount as 0.75 USD in and 2.95 USD out (ZenMux), and Artificial Analysis lists LongCat-2.0 at the same 0.75 and 2.95 (Artificial Analysis).

For scale: an agent session that sends 2 million input tokens, 90% of them from cache, and writes 100,000 tokens costs about 0.19 USD at the launch price (0.06 for uncached input, 0.01 for cached, 0.12 for output). Cheap caching is what makes long agent loops affordable here.

The rest of the billing rules, from LongCat's pay-as-you-go page, Token Pack FAQ and platform FAQ:

• the balance never expires, takes cards and Google Pay, and unused money can be refunded, though API calls are suspended while a refund is reviewed;
• failed requests (401, 403, 429, 500) are not billed;
• Token Packs are fixed 30-day quotas sold in limited flash sales at 10:00, 16:00, 21:00 and 23:00 Beijing time, with cache hits free, and refunded automatically if unused;
• after enterprise verification, monthly pay-as-you-go spending gets 10% off from 75 USD, 20% from 300, 30% from 1,200 and 40% from 6,000.

What the benchmarks show, and who ran them

Meituan's technical blog compares LongCat-2.5 with eight models across 31 tests. A selection, as Meituan published it:

Source: LongCat-2.5 technical blog. Scores with an asterisk come, by Meituan's account, from public reports, blogs or official leaderboards; the others Meituan ran itself and calls indicative only. LongCat-2.5 ran with thinking on, many tests inside Claude Code; BrowseComp used Meituan's own harness with up to 1,024 rounds.

Read plainly, Meituan's own numbers say three things. LongCat-2.5 is far ahead of LongCat-2.0 on long agent work. It sits level with the open Chinese models in the table on Terminal-Bench 2.1 and browsing, and on Terminal-Bench 4.0 it beats all of them except GLM-5.3. It trails Claude Opus 5 on the harder coding tests and DeepSeek V4.1 Flash on several, and its VideoMME score of 85.2 is below the 90.0 Meituan cites for Kimi K3.

The table also leaves things out. Kimi K3 is there, but not the newer Kimi K3.1 Pro, for which Moonshot reports 65.7% on Terminal-Bench 4.0 against GPT-6 Astra and Claude Opus 5.5, two models Meituan does not compare with (Kimi K3.1 Pro guide). And none of it is independent: on November 3 Artificial Analysis had no LongCat-2.5 page and still showed no Intelligence Index for LongCat-2.0, and OpenRouter listed only LongCat-2.0.

What early reviewers and users report

There is little so far beyond the launch numbers. Startup Fortune noted that Meituan published no head-to-head results against the newest Qwen or DeepSeek models and concluded that what is available is "a cheap, capable API and a promise that the weights are coming" (Startup Fortune).

The 什么值得买 column, published on launch day, argues that LongCat-2.5 saves two hand-offs: describing a screenshot or design to the model, and copying code and errors between a chat window and a terminal. Its conclusion is that the model shortens the way from nothing to a page that runs but not the way from running to deliverable, and it points to the gap on Terminal-Bench 4.0 as evidence that real repository work is where LongCat-2.5 is weakest (什么值得买).

The only live measurements come from a reseller. ZenMux, which sells the Preview, showed it on November 3 at a median 2.33 seconds to the first token, 37.3 tokens a second and a 100% success rate over 24 hours (ZenMux). For comparison, Artificial Analysis measured LongCat-2.0 at 74.9 tokens a second. Neither figure is for the final LongCat-2.5, which ZenMux did not list yet. At 37 tokens a second, a 10,000-token answer takes about four and a half minutes.

Privacy, data and content rules

The API comes with its own terms. Developers keep the rights to their inputs and outputs, but grant LongCat a free, worldwide licence to use inputs, outputs and feedback for service optimisation, brand promotion and marketing; anyone who needs to keep data out of training has to arrange an enterprise deal or a separate written agreement. The terms also forbid using outputs to train competing models, make developers run their own content-safety review of inputs and outputs, and send disputes to arbitration in Singapore under Singapore law (API terms).

The API privacy policy, dated June 30, 2026, says servers may be located in Frankfurt, Hong Kong and mainland China and that personal data may be processed outside your country.

The consumer chat is stricter. Its terms describe LongCat as a service for personal learning, research and entertainment, forbid using it or its outputs commercially without Meituan's permission, say it is provided through facilities in the People's Republic of China, and put disputes before a court in Beijing's Chaoyang District (terms of use). Its privacy policy says personal data is stored in Hong Kong and that data, typically de-identified, is used to train and optimise AI models. Business use belongs on the API, not in the chat.

Who should not use LongCat-2.5

• Anyone who needs to run or fine-tune the model on their own hardware now: the weights are not out, and LongCat-2.0, which is, reads text only.
• Teams whose code or customer data must not be used to improve someone else's model and who have no written agreement with Meituan.
• Anyone who needs audio input, speech output or image generation from the same model.
• Buyers who must budget on a fixed price: the only published price is a launch discount with no end date.
• Anyone who wants outside benchmark evidence before choosing: none existed on November 3, 2026.
• Workflows that depend on low temperature or repeatable sampling.
• Anyone planning to use outputs from the LongCat chat commercially, which its terms forbid without Meituan's permission.

Alternatives

Prices from each maker's own price page on November 3, 2026, in USD per million tokens:

Sources: DeepSeek, Z.ai, Moonshot AI, Alibaba Cloud pricing and Qwen-Omni page, Google pricing and Gemini models; weights from Hugging Face (DeepSeek, Z.ai, Moonshot AI).

DeepSeek V4.1 Flash is the better choice for most text coding. At peak hours, 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, it costs exactly LongCat-2.5's launch price, half that the rest of the time, and the same goes for its cached input. Meituan's own table puts it ahead on Terminal-Bench 2.1, DeepSWE and NL2Repo, and its weights are public under MIT. LongCat-2.5 wins if you need video input or the Terminal-Bench 4.0 score matters to you.

GLM-5.3 from Z.ai is the better choice if you want open weights at frontier scale today: it is on Hugging Face and edges LongCat-2.5 on Terminal-Bench 4.0 in Meituan's table. It costs almost five times as much per input token and is not among Z.ai's vision models.

Kimi K3.1 Pro is the better choice when coding quality matters more than price. Moonshot's own figure on Terminal-Bench 4.0, 65.7%, is far above the 38.5 Meituan reports for LongCat-2.5, though the runs are not comparable, and it reads images and video too. It costs about seven times as much per input token and eight times per output token. Moonshot calls it open-weight, but on November 3 only Kimi K3 was on its Hugging Face page.

Qwen3.8-Omni-Flash from Alibaba is the better choice if your agent must hear as well as see: it takes audio alongside text, images and video, has a 1M-token context and costs half as much for input and under 40% as much for output. It answers in text only, as LongCat-2.5 does.

Gemini 4 Argon is the better choice for the broadest input, including audio and PDFs, from Google. It costs nearly seven times as much for input and eight times for output, doubling on December 1, and has no free tier on the API.

If you need LongCat weights today, LongCat-2.0 is on Hugging Face under MIT and on OpenRouter; it shares LongCat-2.5's scale but not its vision. ZenMux resells the Preview.