LongCat-2.5 Review: Cheap Agent Model, Weights Still Pending
Meituan, the Chinese local-services group, released LongCat-2.5 on 2 November: a 1.6-trillion-parameter model for long-running agent tasks, priced from $0.30 per million input tokens. The benchmark placement is Meituan's own, no independent index lists the model yet, and the weights are promised, not published.
Key takeaways
- Meituan released LongCat-2.5 on 2 November 2026 with 1.6 trillion total parameters, about 48 billion active per token, and native image and text input, built for long agent and coding tasks (company announcement; the 48 billion figure is Meituan's, as relayed by IT Home) [1][3].
- Meituan says its scores fall between Claude Opus 4.8 and Claude Opus 5 and beat Opus 5 on some tests. These are company claims; Artificial Analysis showed no LongCat-2.5 entry when we checked on 3 November [1][5].
- The API costs $0.30 per million uncached input tokens, $0.006 cached and $1.20 output, a limited-time discount that Meituan says may change [2].
- GitHub and Hugging Face releases are described as coming soon, with no date or licence. The predecessor, LongCat-2.0, is under the MIT licence [3][4].
What Meituan released
Meituan announced the model on 2 November as a native multimodal system for long-horizon agent tasks [1]. It turns images into discrete tokens that the same backbone processes alongside text, which Meituan says supports visual coding, interactive web pages and 3D tasks [3]. For software work it used large-scale reinforcement learning on long task sequences and turns failures from real agents into training tasks [3].
LongCat-2.5 builds on LongCat-2.0, which Meituan released in June, and keeps the same size [1][4]. It is a mixture-of-experts design, so only about 3% of the parameters (48 billion of 1.6 trillion) work on each token, which is how such models hold serving costs down [3].
Meituan says training ran entirely on domestic Chinese computing hardware and does not name the chip suppliers; the LongCat-2.0 model card refers to "AI ASIC superpods" [1][4]. For inference it quantises weights and activations to 4 bits (W4A4). In a Meituan online stress test, throughput rose about 40%, time to first output fell about 16% and per-token generation time about 20% against an 8-bit baseline. All of these are company figures [1].
Whose benchmarks?
Only Meituan's. It says LongCat-2.5 sits between Claude Opus 4.8 and Opus 5 on several mainstream benchmarks and beats Opus 5 on some [1]. Both are older than Claude Opus 5.5: our 27 October review recorded Artificial Analysis scoring Claude Opus 5.5 at 58 and Kimi K3.1 Pro at 55 [11]. Artificial Analysis's LongCat page, read on 3 November, lists LongCat 2.0 (76 output tokens per second) and a lite model, but no 2.5 [5].
Until an independent index runs it, treat the placement as unverified. Early comments on Hacker News are mixed and anecdotal: some report fast output, while one commenter found it weaker on basic prompts than its charts suggest [10].
| Model | Price per million tokens (input / output) | Artificial Analysis index | Weights |
|---|---|---|---|
| LongCat-2.5 | $0.30 / $1.20, limited-time [2] | none listed yet [5] | promised, no date or licence [3] |
| Kimi K3.1 Pro | $2 / $10 [11] | 55 [11] | due by 4 November, per Moonshot [11] |
| DeepSeek V4.1-Flash-1017 | $0.15 / $0.60 off-peak, double at peak [12] | 45 [12] | MIT licence on Hugging Face [12] |
Meituan's price list also shows ¥2, ¥0.04 and ¥8 per million tokens for uncached input, cached input and output [2]. Cached input costs 2% of the uncached rate, which matters for coding agents that resend long conversations each turn.
Strengths, limits and best fit
Strengths. The per-token price is low next to Kimi K3.1 Pro, though DeepSeek's Flash model is cheaper still [11][12]. It accepts images, and Meituan trained it on sequences of up to 1 million tokens [1].
Limits. There is no independent score, no weights and no licence. The discount is limited-time [2]. Meituan's comparison names older Anthropic models, and its stated aim is to serve its own local-services business, where it has built an AI workbench for merchants [1]. We did not test sign-up, data handling or availability from outside China.
Best fit. Cost-sensitive coding and visual-coding agents whose data may be processed on a Chinese-operated API. Teams that need to self-host, or want independent scores first, should wait.
Open question: weights and licence
Meituan has not said what the GitHub and Hugging Face releases will contain or when they will appear [3]. China's draft AI export catalogue, published on 16 October with comments open to 15 November, is part of the background (see our 17 October report). Neither Meituan nor the ministries has said it affects this release, and we found no source that links the two [13].
The global picture
China, consumers. Apple said on 31 October (China time) that its first Apple Intelligence features reach China in the iOS developer beta from 4 November, with a public release in December [6]. Alibaba's Qwen will sit inside Siri, Writing Tools and Visual Intelligence, while Siri AI and other WWDC 2026 features are excluded [6].
Gulf. Abu Dhabi signed four agreements with OpenAI on 2 November, including a government pilot on OpenAI's GPT models with UAE-based inference infrastructure and a five-year Frontier platform deal with Inception42, reported by The National, not independently confirmed. The report gives no contract values [7].
Russia. A decree signed on 2 November requires private data-centre owners to reserve 10% of capacity for state use, reported by Reuters via The Moscow Times, not independently confirmed [8]. The Kremlin's site lists a decree of that date on temporary measures to support information resources; we have not read its text [9]. Reuters adds that drones struck Yandex and Sberbank facilities in October (see our 1 November report) [8][14].
What to watch next
- LongCat-2.5 weights: the GitHub and Hugging Face release, its licence and its date [3].
- Independent scores: an Artificial Analysis entry for LongCat-2.5 [5].
- Pricing: the end of the limited-time rate [2].
- Kimi K3.1 Pro: the weights Moonshot promised by 4 November [11].
- Apple in China: the developer beta from 4 November and the December release [6].
Our take
Our analysis: the price and the domestic-compute claim are the news; the quality claim is a company chart until an independent index reproduces it. Meituan's chief executive told investors on 28 August that models should support its core business rather than become a "token factory", according to Meituan, so expect tuning for merchant tools first [1]. This is analysis, not financial advice.
FAQ
Is LongCat-2.5 open source? Not yet. Meituan says GitHub and Hugging Face releases are coming, without a date or licence [3]. LongCat-2.0 is under the MIT licence [4].
How much does LongCat-2.5 cost? $0.30 per million uncached input tokens, $0.006 cached and $1.20 output, a limited-time discount that may change [2].
Is it better than Claude Opus 5? Meituan says it beats Opus 5 on some tests. No independent evaluator has confirmed that [1][5].
Where can I try it? On Meituan's LongCat site and API platform [1][2].
Sources
- 美团发布LongCat-2.5大模型, Meituan, 2 November 2026. Primary (company announcement, in Chinese).
- LongCat-2.5 models and pricing, LongCat API Docs, read 3 November 2026. Primary.
- 美团龙猫发布 LongCat-2.5 原生多模态 AI 模型, IT Home, 2 November 2026. Relays Meituan's announcement (in Chinese).
- meituan-longcat/LongCat-2.0, Hugging Face model card, read 3 November 2026. Primary.
- LongCat models, Artificial Analysis, read 3 November 2026. Independent evaluator; primary for its own listings.
- Apple 智能即将登陆中国, Apple Newsroom (China), 31 October 2026. Primary (in Chinese).
- Abu Dhabi agrees OpenAI deals to transform government services and education, The National, 2 November 2026.
- Putin Orders Private Data Centers to Set Aside 10% of Capacity to Support IT Infrastructure, Reuters via The Moscow Times, 2 November 2026.
- Новости и документы, President of Russia, 2 November 2026. Primary (official listing of the decree title; text not read).
- LongCat-2.5: Native Multimodality for Agentic Tasks, Hacker News discussion, read 3 November 2026. Anecdotal user comments only; Tier 3.
- Kimi K3.1 Pro Review: Strong Scores, Slow, Weights Due 4 Nov, AI Worldwire, 27 October 2026.
- DeepSeek V4.1-Flash-1017 Review: Low Price, US Agency Ban, AI Worldwire, 20 October 2026.
- China Drafts AI Export Curbs: Models, Agents, Registration, AI Worldwire, 17 October 2026.
- AI Data Centres Become Targets: Russia and the Gulf Respond, AI Worldwire, 1 November 2026.
Disclosure: the AI system that assisted this article, Claude, is made by Anthropic, whose models Meituan uses as its benchmark reference. AI Worldwire is independent and has no commercial relationship with the companies covered, unless stated. Research and drafting are assisted by AI; every article is reviewed and approved by a human editor before publication. Subscribe to the AI Worldwire newsletter. Last updated: 3 November 2026.