September preview becomes the finished version
Meituan, a major Chinese provider of lifestyle services, released its in-house large language model "LongCat-2.5" in a Nov. 1 post on its official blog. It is the finished version of the preview released on Sept. 25, aimed at "agent" uses in which AI carries out multistep work such as development tasks on its own. Its size is the same as the preview's at 1.6 trillion parameters, of which about 48 billion are used in each pass. The amount it can read at once (context) is 1 million tokens, and it can output up to 128,000 tokens at a time.
According to Meituan, the model is designed to handle images and text in the same architecture, and also supports writing programs from the look of a screen, design and 3D creation. The company said it expanded reinforcement learning aimed at long tasks and shored up weaknesses by turning tasks that real agents failed at into training material. It set the goal of completing entire software projects rather than code fragments.
The release was reported in the early hours of Nov. 2 Japan time by an X account that automatically detects and reports AI model releases.
Meituan had released the preview on its official X account on Sept. 25 (the evening of the same day Japan time). No benchmark results were released for the preview.
LongCat-2.5-Preview is now live. 1.6T parameters. ~48B active. A 1M-token context window. Natively multimodal. Built to take on long-horizon tasks. From terminals and browsers to GUIs, spreadsheets, and design tools.
API input at $0.30, the same as DeepSeek's small model
Developer API pricing is a limited-time discount of $0.30 for input and $1.20 for output per million tokens, the units of text the AI processes. Input when the same text is read repeatedly (cached input) costs $0.006. No end date for the discount was given.
OpenAI's "GPT-6.1 Terra" costs $1 for input and $5 for output, and Anthropic's "Claude Sonnet 5.5" costs $2 for input and $10 for output, so LongCat-2.5 undercuts both. It matches China's DeepSeek's "DeepSeek V4.1 Flash" on both input and output at peak-hour rates. DeepSeek charges half that at other times.
- Input
- Output
Tops Opus 5 on web research tasks in own evaluation
On its official blog, Meituan published results comparing the model with models from OpenAI, Anthropic, China's Zhipu (GLM), Moonshot AI and DeepSeek on 31 benchmarks. On "BrowseComp," which tests finding answers by researching the web, it scored 91.9, above Anthropic's "Claude Opus 5" at 90.8 and OpenAI's "GPT-5.6-Sol" at 90.4, the highest among the five models in the table. On "Terminal Bench v2.1," which tests solving command-line tasks, it scored 88.4, close to Claude Opus 5's 89.1 and GPT-5.6-Sol's 88.8.
15 points behind Opus 5 on a newer software development benchmark
On the newer "v4.0" version of Terminal Bench, however, it scored 38.5, 15.4 points below Claude Opus 5's 53.9. On "DeepSWE v1.1," which tests solving software development tasks, it scored 64.6, below all four other models in Meituan's comparison table (66.9 to 72.7). Compared with the previous model, LongCat-2.0, the Terminal Bench v4.0 score rose from 3.0 and the DeepSWE v1.1 score from 9.7.
Comparisons against the previous generation, no third-party evaluation yet
None of the U.S. models Meituan compared against are the latest. Anthropic released "Claude Opus 5.5" on Sept. 23 and OpenAI released "GPT-6.1 Sol" on Sept. 29, but neither is in the table.
As of Nov. 3, U.S. evaluation company Artificial Analysis had not added LongCat-2.5 to its rankings. There are not yet any third-party figures comparing it with other companies' models under the same conditions.
DeepSeek V4.1 Flash, which has the same pricing, scored 74.2 on DeepSWE v1.1 according to DeepSeek's announcement. However, the two companies used different measurement environments and reasoning settings, so the figures cannot be compared directly. V4.1 Flash's model weights are public, while Meituan has said only that LongCat-2.5's weights are "coming soon" on GitHub and Hugging Face.
| Benchmark | LongCat-2.5 | Claude Opus 5 | GPT-5.6-Sol | GLM-5.3 | Kimi K3 |
|---|---|---|---|---|---|
| Terminal Bench v2.1 | 88.4 | 89.1* | 88.8* | 88.2* | 88.3* |
| Terminal Bench v4.0 | 38.5 | 53.9* | 37.3* | 41.8* | 12.6* |
| DeepSWE v1.1 | 64.6 | 68.8* | 72.7* | 66.9* | 67.5* |
| BrowseComp | 91.9 | 90.8* | 90.4* | 81.0 | 91.2* |
| HLE w/tools | 63.1 | 64.7* | 58.0* | 62.5* | 56.0* |
| JobBench | 65.4 | 66.7* | 45.4* | 61.4* | 52.9* |
| MCP Atlas | 83.1 | 85.8* | 81.8* | 84.2* | 84.2* |
| 3DCodeBench Text to 3D | 55.4 | 54.3 | 52.3 | 53.4 | 49.3 |
| Vision2Web | 78.6 | 80.2* | 73.5* | 77.8* | 79.0* |
Values marked * were taken by Meituan from published reports and leaderboards, and unmarked values were measured by Meituan, which calls them "reference values." LongCat-2.5 was measured in a setting that outputs its reasoning process (thinking mode). Some coding items were measured on Anthropic's AI coding tool "Claude Code."
Trained and served on Chinese-made chips
Meituan said it carried out both training and inference on Chinese-made AI chips. It split processing so that long texts could be trained on chips with less than 80GB of memory each, and for inference it runs the model with its numbers compressed to 4 bits. Compared with running at 8 bits, throughput for reading text rose 43.3% and throughput for generating text rose 42.77%, the company said. It did not disclose the chipmaker.
At its April-June earnings call in August, Meituan described its previous model, LongCat 2.0, as one of the first trillion-parameter-class models trained solely on Chinese infrastructure. Revenue for the quarter rose 14.4% from a year earlier to 104.6 billion yuan, and adjusted net profit returned to the black at 2.5 billion yuan. Research and development spending rose to 7.3% of revenue, which the company attributed to increased investment in AI. CEO Wang Xing said Meituan does not intend to compete as a factory mass-producing tokens, adding that AI investment would be judged on its return.
- LongCat 2.0 released, with plan to release weightsReport
- Earnings call describes LongCat 2.0 as trained solely on Chinese infrastructureOfficial
- LongCat-2.5-Preview released on API, with no benchmarksOfficial
- LongCat-2.5 released, with 31 benchmarks publishedOfficial
Offered in chat service and API, works with Claude Code and Codex
LongCat-2.5 is available in Meituan's chat service and its developer API. The API supports both OpenAI and Anthropic formats, and Meituan says it can be used in the AI coding tools "Claude Code," "Codex" and "Hermes." According to APIYI, to mark the release, new users are being given 10 million tokens and existing users 5 million tokens for free.
