Meituan LongCat-2.5 (1.6T MoE, Multimodal)
Meituan's long-horizon agent model, trained entirely on Chinese domestic accelerators. Adds native vision to the LongCat-2.0 line, with self-reported coding-agent scores between Claude Opus 4.8 and Opus 5 at a very low limited-time price.
Interactive Token Cost Estimator (LongCat-2.5)
Architectural Strengths & Target Workloads
- Vendor-reported 88.4 on Terminal Bench v2.1, 38.5 on v4.0, 64.6 on DeepSWE v1.1 and 83.1 on MCP Atlas
- 1M-token context with up to 128K output tokens; cached input at $0.006 per 1M
- Trained across multiple agent harnesses; works with Claude Code, Codex and Hermes via Anthropic/OpenAI-compatible APIs
Primary Engineering Workloads: Long-Horizon Coding Agents · Visual / 3D Coding · Low-Cost Agentic Batch Work.
Frequently Asked Questions (LongCat-2.5)
How much does Meituan LongCat-2.5 (1.6T MoE, Multimodal) cost per 1M tokens?
Meituan LongCat-2.5 (1.6T MoE, Multimodal) is priced at $0.30 per 1M uncached input tokens, $0.01 per 1M cached input tokens, and $1.20 per 1M output tokens (LongCat API, limited-time discounted pricing. OpenAI-compatible (api.longcat.chat/openai) and Anthropic-compatible (api.longcat.chat/anthropic) endpoints.).
What is the context window of Meituan LongCat-2.5 (1.6T MoE, Multimodal)?
It supports up to 1,048,576 tokens (1M) (1,048,576 tokens) and a maximum generation output of 131,072 tokens.
Meituan's LongCat-2.5: A 1.6T-Parameter Multimodal Agent Model at $0.30 / $1.20 per Million Tokens
Trained entirely on Chinese domestic accelerators, LongCat-2.5 adds native vision to Meituan's trillion-parameter line. Its self-reported coding-agent scores land between Claude Opus 4.8 and Opus 5, with a 1M-token context and an Anthropic-compatible API.
Read Full Technical Dispatch (5 min read) →