Meituan's LongCat-2.5 Drops Into Claude Code at $0.30/$1.20. Its Setup Caps You at 200K

• Book: System Design Pocket Guide: Fundamentals
• The series: Pocket Guides for Developers — 8 short books on agents, RAG, prompts, observability, events, databases and system design — the whole series
• My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
• Me: xgabriel.com | GitHub

Meituan released LongCat-2.5 on November 2: a 1.6-trillion-parameter model with about 48B active per token, according to TechNode. Its API docs list a 1M-token context window, an Anthropic-compatible endpoint, and a guide for running it inside Claude Code. The launch price is $0.30 per million input tokens and $1.20 per million output.

Follow that guide as written and Claude Code will compact your conversation at 200K tokens. Claude Code does this to any model name it doesn't recognize, and the fix is one environment variable.

What Meituan shipped on November 2

Meituan's announcement describes a native multimodal model aimed at long-horizon agent tasks, trained entirely on domestic Chinese chips, with W4A4 quantization for inference. TechNode reports that Meituan turns failures real agents hit into reproducible training tasks, and that text and images go through one token framework.

On quality, Meituan's own summary says its scores on several mainstream benchmarks fall between Claude Opus 4.8 and Claude Opus 5, and beat Opus 5 on some (announcement). Look at what it's compared against. A commenter in the Hacker News thread quotes the comparison set from Meituan's page: GPT-5.5, GPT-5.6 Sol, Opus 4.8 and Opus 5. In Claude Code, the opus alias has resolved to Opus 5.5 since v2.1.280, so the Opus you'd actually be replacing isn't on the chart. Meituan's are the only scores published so far.

Weights aren't out. TechNode says the GitHub and Hugging Face releases are "planned to follow", with no date and no word on whether they include weights. The predecessor's weights did ship: LongCat-2.0 is on Hugging Face under the MIT License, at the same 1.6T total and about 48B active. One HN commenter summed up the 2.5 page: "Is this supposed to be downloadable weights? Open weights? Who knows?"

Why the Claude Code guide gets you 200K

LongCat's Claude Code guide has you put an env block in ~/.claude/settings.json that points ANTHROPIC_BASE_URL at https://api.longcat.chat/anthropic and sets ANTHROPIC_MODEL, the Opus and Sonnet defaults and the old small-fast variable to LongCat-2.5. It never sets ANTHROPIC_DEFAULT_HAIKU_MODEL.

Claude Code has never heard of LongCat-2.5. Its model configuration docs say that when it doesn't recognize a session's model ID, "it can't tell the model's context window and assumes 200K tokens", and it auto-compacts at that assumed size. The same page gives the fix for any other window: set CLAUDE_CODE_MAX_CONTEXT_TOKENS. For an unrecognized ID that doesn't start with claude-, which LongCat-2.5 doesn't, the variable "takes effect on its own."

A 1M-token model, a 200K assumption. Claude Code compacts at the window it thinks the model has.

The guide sets three other variables you should check:

• CLAUDE_CODE_MAX_OUTPUT_TOKENS is set to 131072. For a model ID Claude Code can't resolve, the env-vars reference says the cap is 128000, and that raising this value "reduces the effective context window available before auto-compaction triggers."
• ANTHROPIC_SMALL_FAST_MODEL is set, and Claude Code's docs mark it deprecated in favor of ANTHROPIC_DEFAULT_HAIKU_MODEL, which also drives background work.
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC is set. Per the reference, that turns off auto-updates, telemetry and feature-flag fetching, and setting it to 0 "still disables this traffic." So Claude Code stops updating itself. To turn it back on, unset it. The wrapper below leaves it out.

A setup that leaves your other sessions on Claude

The bigger issue is where the guide puts all this. Claude Code's settings docs say ~/.claude/settings.json applies to "you, in every project on this machine." Paste the guide there and every Claude Code session you start goes to LongCat.

If you want to try it next to your normal setup, scope it to one launch with a wrapper script:

#!/usr/bin/env bash
# ~/bin/claude-longcat
export ANTHROPIC_BASE_URL="https://api.longcat.chat/anthropic"
export ANTHROPIC_AUTH_TOKEN="$LONGCAT_API_KEY"
export ANTHROPIC_MODEL="LongCat-2.5"
export ANTHROPIC_DEFAULT_OPUS_MODEL="LongCat-2.5"
export ANTHROPIC_DEFAULT_SONNET_MODEL="LongCat-2.5"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="LongCat-2.5"
export CLAUDE_CODE_SUBAGENT_MODEL="LongCat-2.5"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000
exec claude "$@"

chmod +x it, export LONGCAT_API_KEY in your shell, and run claude-longcat in a repo. Plain claude stays on your Anthropic account. CLAUDE_CODE_SUBAGENT_MODEL sets the default for subagents that aren't given a model another way. I haven't run LongCat-2.5 myself, so check /status and /context on your first session before you trust the window.

What it costs depends on your cache

The pricing page labels these a limited-time discounted price and says "Prices are subject to change":

Cached input is one fiftieth of uncached. Coding agents resend the whole conversation every turn, so the cache hit rate decides the bill. On the listed rates, a turn that sends 100K input tokens costs $0.03 uncached and $0.0006 if it all hits the cache. Output usually dominates once the cache is warm.

Same tokens, two prices. On an agent loop, most of what you resend should land in the small stack.

If you call it from your own code, rate-limited requests come back as HTTP 429 with a retry_after field, and the quick start recommends exponential backoff. The chat API reference says temperature is "Fixed at 1.0 with thinking enabled and 0.7 with thinking disabled", with other values "overridden." If your eval harness sets temperature 0 for repeatable runs, it won't get it here.

Running it next to Claude Code in Hermes

I build Hermes IDE, so weigh this section with that in mind. Hermes runs Claude Code, Codex, Copilot, OpenCode and the rest side by side, each in its own terminal. ⌘N (Ctrl+Shift+N on Windows and Linux) starts each task on its own hermes/<name> branch in its own worktree, so you can give stock Claude Code and the LongCat wrapper the same ticket without either touching the other's files.

The ⌘N agent picker in Hermes 2.3.4: Claude Code and Codex installed, the other supported agents listed, and Custom agent at the end.

The wrapper goes in as a Custom agent with claude-longcat as its command. Whatever spend figure you see for that session, check it against LongCat's own billing. Hermes is free on macOS, Windows and Linux, and it isn't affiliated with Meituan, Anthropic or Nous Research.

Before you move real work to it

One HN commenter called it "benchmaxxed" after a few prompts. That's one person's gut read. Give Meituan's charts the same caution until someone else runs the benchmarks.

There's also a policy question nobody has answered. China's Ministry of Commerce and Ministry of Science and Technology published a draft export catalogue on October 16, open for comment until November 15. A Cloud Security Alliance note reads it as requiring reporting and automatic licensing before newly developed restricted model technology is open-sourced, while weights already public stay outside it. Meituan hasn't said whether LongCat-2.5 counts, and neither have the ministries.

The API route works today and costs little to try. Would you put a model whose only scores are its maker's on a real ticket, or wait for the weights and an independent run? Tell me in the comments.

If this was useful

Most of this post was about two old system design problems: a cache that decides what a request costs, and a rate limit that tells you when to back off. The System Design Pocket Guide: Fundamentals covers caching, rate limiting and the other building blocks an agent backend leans on. It's one of 8 short books in the Pocket Guides series, on agents, RAG, prompts, observability, events, databases and system design.

If you want to race a new model against Claude Code on the same repo, download Hermes IDE (free, source-available under BSL 1.1) or star it on GitHub.