What is LongCat-2.5 and why does a 1-million-token context matter?

Meituan released LongCat-2.5 on November 2: a 1.6-trillion-parameter mixture-of-experts model with about 48 billion active parameters, multimodal, a 1-million-token context and 128K max output, priced at $0.30/$1.20 per million tokens with OpenAI- and Anthropic-compatible APIs. Why long context plus cheap cached input changes long agent jobs — and how to evaluate it safely.

Short answer: LongCat-2.5 is a new AI model from Meituan — the Chinese food-delivery and local-services company — released on November 2, 2026: a mixture-of-experts model with 1.6 trillion total parameters but only about 48 billion active per token, natively multimodal, with a 1-million-token context and outputs up to 128,000 tokens. It is priced at $0.30 per million input tokens and $1.20 per million output tokens on Meituan’s platform, with cached input far cheaper. The weights are not open yet.Why it matters: context is an AI’s working memory: a 1-million-token window lets an agent keep an entire codebase, a long contract set or months of notes in view at once, instead of forgetting the beginning by the end.

Pair that with cheap cached input — the instructions and files an agent re-reads on every step — and multi-hour work that used to be priced out becomes practical: overnight code refactors, full-document reviews, research runs that check their own work. The caveat: a bigger window is not better judgment, and the benchmark claims so far are Meituan’s own.How to evaluate it safely, step by step:1. Use the API formats you already have: LongCat speaks both OpenAI and Anthropic request formats, so most tools connect by changing a base URL and a key.2.

Test on copies, never originals: run it against a duplicate of a real project or document set, keeping your current model’s output as the control.3. Score the finished job, not the demo: judge the final output at the end of a long run — that is where long-context models quietly drop early details.4. Watch the cached-input line on the bill: long agent jobs re-send huge prompts repeatedly, so caching is where the saving lives — confirm you are actually getting cache hits.5.

Keep secrets out: with the weights not yet open, everything you send runs on Meituan’s servers — treat it like anything pasted into a third-party website.Pro tip: run the same long task on both models and compare the finished work before the prices — cheap tokens are only cheap if the job needs no redoing.Ad’s take: A million tokens of memory at these prices is a change you feel over a month, not in a headline — agents that check their own work for hours just got far more affordable.