LongCat-2.5 Guide: API, Pricing, Vision and Claude Code

Last updated: November 3, 2026

LongCat-2.5 is Meituan’s new 1.6-trillion-parameter mixture-of-experts model for agentic coding, announced November 1, 2026. It adds native image and video understanding, a 1M-token context window and 256K output, works in Claude Code and Codex, and costs $0.30 input and $1.20 output per million tokens at a limited-time discount.

Contents

What’s new at a glance
What changed from LongCat-2.0
API access, endpoints and limits
Pricing and billing
Image and video input
Using it in Claude Code and Codex
Benchmarks and comparison
Known limitations
Frequently asked questions

Meituan’s LongCat team published the LongCat-2.5 technical blog on November 1, 2026, and the LongCat API change log lists general availability on November 2. A LongCat-2.5-Preview with image understanding had appeared on September 25. This guide covers what changed, how to call it, what it costs and its limits.

What’s new at a glance

Item	LongCat-2.5
Model name	`LongCat-2.5` on OpenAI- and Anthropic-compatible endpoints
Size	1.6 trillion total parameters, about 48 billion activated per token
Context and output	1,048,576 tokens of context; `max_tokens` up to 262,144
Input types	Text, images (Base64 or URL) and video (URL only)
Reasoning control	`"thinking": {"type": "enabled"}` or `"disabled"`; `thinking_budget` up to 262,144
New endpoint	OpenAI Responses-compatible `POST /openai/v1/responses`, used by Codex
Price (limited-time)	$0.30 uncached input, $0.006 cached input, $1.20 output per 1M tokens
Open weights	Listed as coming soon; no date published

What changed from LongCat-2.0

LongCat-2.5 keeps the size of LongCat-2.0 but adds two things. First, native multimodality: the model turns images into multi-level discrete tokens, so the same backbone reads visual and text tokens in one stream. Meituan says this supports visual coding, creative design and 3D tasks.

Second, more long-horizon agent training. The blog describes large-scale reinforcement learning on tasks built from GitHub pull requests and internal coding work, with failed real-world agent runs converted into verifiable training tasks. It also adds penalties for repetitive reasoning and repeated tool calls.

The change log adds Codex to the list of supported harnesses, alongside Claude Code and Hermes. The preview listed Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code.

API access, endpoints and limits

Per the LongCat API Quick Start, one API key works on both base URLs: https://api.longcat.chat/openai for Chat Completions and https://api.longcat.chat/anthropic for the Messages format. The Responses API page adds reasoning effort levels from none to xhigh, parallel tool calls (on by default), custom tools and JSON schema output.

The Model Parameters page sets the rules power users need to know: temperature is fixed at 1.0 with thinking on and 0.7 with thinking off, top_p defaults to 0.95, top_k ranges from -1 to 100, and repetition_penalty ranges from 1 to 1.5. Out-of-range values return a 400 error. Rate limits apply per API key and return HTTP 429 with a retry_after value; the docs do not publish numeric limits.

To make a first call:

Sign in to the LongCat API Platform and create a key on the API Keys page. The key is shown only once.
Top up under Billing, then Recharge. Payment options include Alipay, WeChat Pay, credit or debit cards and Google Pay.
Point the OpenAI SDK at https://api.longcat.chat/openai or the Anthropic SDK at https://api.longcat.chat/anthropic.
Set model to LongCat-2.5 and add "thinking": {"type": "enabled"} for harder tasks.
Read reasoning_content and usage.completion_tokens_details.reasoning_tokens in the response to track thinking cost.

Pricing and billing

The pay-as-you-go page labels every price as a limited-time discount and does not give an end date. LongCat’s own worked example, with 600,000 uncached input tokens, 400,000 cached input tokens and 500,000 output tokens, costs $0.7824.

Item	USD per 1M tokens	CNY per 1M tokens
Uncached input	$0.30	¥2
Cached input	$0.006	¥0.04
Output	$1.20	¥8

Failed requests (401, 403, 429, 500) are not billed, prepaid balance does not expire, and Token Packs (a fixed quota valid for 30 days) are used before the pay-as-you-go balance. The LongCat FAQ lists enterprise discounts from 10% at $75 of monthly spend to 40% at $6,000 for verified organizations.

Image and video input

The Visual Understanding docs show images sent as image_url blocks (public URL or Base64 data URL) and videos as video_url blocks (public URL only). The Anthropic endpoint accepts image blocks with a base64 or url source.

Images: PNG, JPG, JPEG, WebP or GIF, up to 10 MB each and 100 per request. SVG and HEIC are not supported, and an aspect ratio over 200:1 returns an error.
Videos: up to 100 MB each, under 3,600 seconds, and 10 per request.

Using it in Claude Code and Codex

The Claude Code setup guide puts the key and model in ~/.claude/settings.json: set ANTHROPIC_AUTH_TOKEN, ANTHROPIC_BASE_URL to the Anthropic endpoint, and ANTHROPIC_MODEL, ANTHROPIC_SMALL_FAST_MODEL and the default Sonnet and Opus slots to LongCat-2.5, with CLAUDE_CODE_MAX_OUTPUT_TOKENS at 262144. Because every slot points at one model, background calls cost the same as main turns. For plan details on Anthropic’s side, see the Claude Code power-user guide.

For Codex, the Codex setup guide sets model = "LongCat-2.5", model_reasoning_effort = "high", base_url = "https://api.longcat.chat/openai/v1" and wire_api = "responses" in ~/.codex/config.toml, with the key stored as OPENAI_API_KEY in ~/.codex/auth.json. Our Codex CLI 0.166 guide covers the recent Codex features, and the OpenCode 2.2 guide covers another harness LongCat supports.

A practical tip: keep a stable prompt prefix so more of each turn bills as cached input, which costs 2% of the uncached rate.

Benchmarks and comparison

All figures below come from Meituan’s own evaluation table, run with thinking on. Meituan says competitor scores marked as self-run are indicative only, and no independent results were available when this guide was written.

Benchmark	LongCat-2.5	LongCat-2.0	DeepSeek V4.1 Flash	Claude Opus 5
Terminal Bench v4.0	38.5	3.0	31.2	53.9
DeepSWE v1.1	64.6	9.7	74.2	68.8
NL2Repo	60.5	40.9	65.4	75.3
BrowseComp	91.9	79.9	91.5	90.8
MCP Atlas	83.1	61.8	84.4	85.8

The jump over LongCat-2.0 is large on long-horizon tasks. Against the same-price rival in that table, results are mixed: LongCat-2.5 leads on Terminal Bench v4.0 and BrowseComp, while DeepSeek V4.1 Flash leads on DeepSWE and NL2Repo. Compare with our DeepSeek V4.1 Flash guide, the Kimi K3.1 Pro guide and the MiniMax M3.1 Flash guide for other low-cost agent models.

Known limitations

No open weights yet. Cocoloop reported that the GitHub and Hugging Face links read coming soon, unlike LongCat-2.0, which was open-sourced at launch.
Prices are promotional with no stated end date.
Temperature cannot be tuned, which affects sampling sweeps in evaluation harnesses.
Numeric rate limits are not published.
Cocoloop noted that the comparison set excludes the GPT-6 series and Opus 5.5, and no independent replication exists yet.

Frequently asked questions

What is LongCat-2.5?

LongCat-2.5 is Meituan’s native multimodal model for long-horizon agent tasks, announced November 1, 2026. It has 1.6 trillion total parameters with about 48 billion active per token, accepts text, images and video, and is available through the LongCat API Platform under the model name LongCat-2.5.

How much does LongCat-2.5 cost?

The LongCat pay-as-you-go page lists a limited-time discounted price of $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens (¥2, ¥0.04 and ¥8 on the China site). Only successful HTTP 200 requests are billed.

Is LongCat-2.5 open source?

Not yet. Cocoloop reported that the GitHub and Hugging Face links on the release page read coming soon, with no date or license announced. LongCat-2.0 was open-sourced at launch. For now, LongCat-2.5 is available through the API and LongCat’s website.

What is the LongCat-2.5 context window?

The Model Parameters page lists a context length of 1,048,576 tokens (input plus output) and a max_tokens limit of 262,144 tokens per request. The Quick Start describes this as a 1M token context window with up to 256K output tokens.

Can I use LongCat-2.5 in Claude Code or Codex?

Yes. LongCat publishes setup guides for both. Claude Code uses the Anthropic-compatible base URL https://api.longcat.chat/anthropic with ANTHROPIC_MODEL set to LongCat-2.5. Codex uses https://api.longcat.chat/openai/v1 with wire_api set to responses.

Can I change the temperature for LongCat-2.5?

No. Temperature is fixed at 1.0 when thinking is on and 0.7 when thinking is off. Other values are not rejected, but the platform overrides them. You can still adjust top_p, top_k, repetition_penalty and thinking_budget within documented ranges.

Sources

Image: LongCat logo by Meituan, via Wikimedia Commons, public domain (trademark of Meituan), placed on a white background and resized. FreeAnswers is not affiliated with Meituan.

LongCat-2.5: Power-User Guide to the API, Pricing, Vision and Agent Setup