LongCat-2.5: Meituan's 1.6T agent model, drop-in via Claude API
48B active of 1.6T, a 1M-token window, $0.30/$1.20 per Mtok, and self-reported wins over GPT-5.6-Sol — but the open weights aren't out yet.
Copy markdown
Point Claude Code at it today
LongCat-2.5 serves from OpenAI- and Anthropic-compatible endpoints (api.longcat.ai/openai and /anthropic, model name `LongCat-2.5`), so you swap one base URL in Claude Code, Codex, OpenCode, or the OpenAI/Anthropic SDKs and keep your setup. It ships with a 1M-token context, up to 128K output, and launch pricing of $0.30/Mtok in (cache miss), $0.006 cache hit, $1.20 out — cheap for a frontier-class agent model.
Built for long agents, not chat
The 1.6T-parameter MoE activates ~48B per token and is natively multimodal — text and vision in one token stream. Meituan trained it with large-scale RL on long task sequences, including reproductions of real agent failures, and pitches it at whole-project work across terminals, browsers, GUIs, and 3D rather than one-shot prompting.
The benchmarks — mind the asterisk
Meituan's own numbers put it ahead of GPT-5.6-Sol on 20 of 27 shared tests (BrowseComp 91.9, MCP Atlas 83.1, DeepSWE 64.6), but ahead of Claude Opus 5 on only 10 of 27 — and it trails Opus 5 on Terminal-Bench v4.0 (38.5 vs 53.9). Many rival scores were run by Meituan, so treat the lead as indicative until third parties verify.
Open weights 'coming soon'
There's no Hugging Face or GitHub drop yet — both links are marked 'coming soon,' so the API is the only way in for now. Meituan says training and inference were tuned for domestic Chinese accelerators under 80GB each: a sovereignty signal if you want an open 1.6T model that doesn't assume Nvidia.