# LongCat 2.5 Explained

> Meituan's native multimodal model for long-horizon agent work (November 1, 2026): 1.6 trillion total parameters, about 48 billion active per token, 1 million token context, 128,000 token maximum output. API cost calculator at discounted and reference prices, vendor-reported benchmark table with conditions, access checker, and setup steps for the LongCat API. Weights are not released; the API and web chat are the only ways to use it as of November 3, 2026. Independent explainer, not affiliated with Meituan. Updated 2026-11-03. Full page: https://lab.awaara.cn/longcat-2-5/

## What it does

- Cost calculator: enter input tokens per run, cache hit percent, output tokens per run, and runs per day. It splits input into cached and uncached parts, applies the selected price set, and shows cost per run, per day, and per 30 days, with a line by line breakdown. Presets: 1M token batch (800,000 input, 200,000 output, no cache, $0.48 at the discount), agent day (300,000 input, 70 percent cache, 15,000 output, 20 runs), short chat reply.
- Access checker: pick what you need (web chat, API, Claude Code, weights, aggregator, local run) and get the status Meituan published on November 1, 2026. Weights and local runs are not available; the web chat and API are.
- Benchmark explorer: 13 rows from Meituan's 31 row official table, filterable by code agent, general agent, and visual groups. Comparison scores marked with an asterisk come from public reports or leaderboards; other comparison scores are Meituan self-run and indicative only.
- Setup steps: OpenAI and Anthropic compatible endpoints on api.longcat.chat, model name LongCat-2.5, and Claude Code variables. Everything on the page runs locally in the browser. Nothing is sent to a server.

## Facts

- Release: November 1, 2026, blog post "Introducing LongCat-2.5". Preview reached the API on September 25, 2026, with no benchmark table. LongCat 2.0 launched June 30, 2026.
- Scale: 1.6 trillion total parameters, about 48 billion activated per token, as reported by Meituan. The figures were disputed before release because they matched the LongCat 2.0 card; the November 1 post states them for 2.5 directly. Cite them as reported by Meituan.
- Context and output: 1,048,576 token context window, 128,000 token maximum output, per the LongCat API docs. Some third party pages list 256,000 or 262,144 maximum output for the preview; the current docs say 128,000.
- Multimodality: images become multi-level discrete tokens and are read by the same backbone as text. Stages: native resolution tokenization (semantic aligned encoder plus residual vector quantization), offline visual token preprocessing, visual embeddings learned jointly with the backbone.
- Training: more long-horizon trajectories in long-sequence training, large-scale long-horizon reinforcement learning, real agent failure trajectories turned into reproducible tasks, penalties for repeated reasoning and repeated tool calls. Expert groups for agent, reasoning, and interaction retained from LongCat 2.0.
- Inference: W4A4 quantization (4 bit weights and activations) on domestic AI accelerators. Meituan reports, against a W8A8 INT8 baseline in online stress tests: prefill throughput up 43.3 percent, decode throughput up 42.77 percent, time to first token down 15.91 percent, time per output token down 19.72 percent.
- Pricing, limited-time discount: $0.30 per million uncached input tokens, $0.006 per million cached input tokens, $1.20 per million output tokens. Meituan has not published a list price or an end date for the discount.
- Reference price used on this page for planning only: $0.75 input, $0.015 cached input, $2.95 output. Borrowed from third party listings for LongCat 2.0. Not an official list price.
- Billing reports (confirm on your account): APIYI reports only HTTP 200 requests are billed; 401, 403, 429, and 500 requests are not billed; prepaid balance does not expire; promotional credits are used first. Rate limit: HTTP 429, code rate_limit_exceeded, retry_after 60, exponential backoff recommended. No RPM or TPM quota is published.
- Benchmark conditions: all LongCat 2.5 scores with thinking on, temperature 1.0, top_p 0.95. Terminal Bench in Claude Code (512K limit), DeepSWE in mini-swe-agent (1M limit), BrowseComp in Meituan's own harness (search and browse tools only, 1,024 rounds cap). No independent reproduction published as of November 3, 2026.
- Benchmark highlights (vendor-reported): BrowseComp 91.9, 3DCodeBench text to 3D 55.4, Terminal Bench v2.1 88.4, Terminal Bench v4.0 38.5 (Opus 5: 53.9*), DeepSWE v1.1 64.6 (DeepSeek V4.1 Flash: 74.2*), JobBench 65.4, Vision2Web 78.6. Gains over LongCat 2.0: Terminal Bench v4.0 3.0 to 38.5, DeepSWE 9.7 to 64.6, AutomationBench 15.8 to 51.8.
- Weights and license: GitHub and Hugging Face buttons on the release post say Coming soon as of November 3, 2026. No license is published for 2.5. LongCat 2.0 shipped MIT licensed weights on July 5, 2026; that does not prove the terms for 2.5.
- Endpoints (from the LongCat API docs): OpenAI format https://api.longcat.chat/openai, Anthropic format https://api.longcat.chat/anthropic. Model name: LongCat-2.5. Older guides list api.longcat.ai; the docs domain is api.longcat.chat.

## Sources

- LongCat official blog, Introducing LongCat-2.5: https://longcat.ai/blog/longcat-2.5
- LongCat API docs, Quick Start: https://longcat.ai/platform/docs
- APIYI, LongCat 2.5 guide: https://help.apiyi.com/en/longcat-2-5-native-multimodal-agent-model-guide-en.html
- TokRace, LongCat 2.5 review: https://www.tokrace.com/en/guides/longcat-2-5-release
- AI Affairs, Meituan releases LongCat-2.5: https://www.aiaffairs.com/technology/meituan-releases-longcat-2-5-long-running-coding-tasks/
- Agmazon, LongCat-2.5-Preview explained: https://agmazon.com/blog/articles/technology/202610/longcat-2-5-preview-complete-guide-en.html
- ExplainX, LongCat 2.5 explainer: https://www.explainx.ai/blog/longcat-2-5-1-6-trillion-agent-model-2026