LongCat 2.5 explained

API cost calculator

LongCat 2.5 bills three token counts separately. Cached input is far cheaper than uncached input, so long agent sessions that re-read the same repository cost much less per run than the headline input price suggests. Enter your own numbers, or start from a preset.

Can I use it today?

Choose what you need. The answer follows the status Meituan published on November 1, 2026, not a guess about what might arrive next.

What LongCat 2.5 is

LongCat 2.5 is a mixture of experts model. Most of its 1.6 trillion parameters stay inactive on any one token. About 48 billion activate per token, according to the official blog post dated November 1, 2026. That design is why a model this large can be priced near $1 per million output tokens on a discount, while a dense model of the same size could not be served at that price.

Native multimodality is the main technical claim. Images are turned into discrete tokens at several levels of detail, and the same backbone reads text and image tokens together. There is no separate vision encoder bolted on afterwards. Meituan says this helps visual coding: reading a screenshot or a design mockup and writing the code that reproduces it, plus creative design and 3D work. The blog post describes three stages: native resolution tokenization with a semantic aligned encoder and residual vector quantization, offline precomputation of visual token ids once the tokenizer is fixed, and visual embeddings learned jointly with the backbone.

The agent claim rests on training, not only on size. Meituan says it added more long-horizon trajectories to long-sequence training, then ran large scale long-horizon reinforcement learning on end to end engineering tasks: working in a complex codebase and building a complete project. It also turned real agent failure trajectories into reproducible, verifiable training tasks, and added penalties for repeated reasoning and repeated tool calls. The point is practical. In a long agent session, the expensive failure is a model that calls the same tool again and again without progress.

Scale, context, and output figures for 2.5 come from the official blog post and the LongCat API docs. The 2.0 weights date and license come from secondary coverage of the 2.0 release. Do not read the 2.0 license across to 2.5.

Meituan also reports inference work that matters for price. Expert weights and activations are quantized to 4 bits (W4A4) and run on native INT4 units on domestic accelerators. In online stress tests against a W8A8 INT8 baseline, Meituan reports prefill throughput up 43.3 percent, decode throughput up 42.77 percent, time to first token down 15.91 percent, and time per output token down 19.72 percent. These are vendor figures from the same blog post. They explain the discount. They do not prove the same speed on your workload.

Benchmarks, with the conditions attached

The official blog post publishes a 31 row table across code agent, general agent, and visual groups. The table below shows 13 rows that cover each group and the comparisons developers ask about most. Every number is Meituan's own measurement unless marked with an asterisk. Read the notes under the table before you rank models on these numbers.

Read the table as a shape, not a ranking. LongCat 2.5 leads this selection on BrowseComp (91.9) and on 3DCodeBench text to 3D (55.4). It trails by a wide margin on the harder terminal test, Terminal Bench v4.0 (38.5 against 53.9 for Opus 5), and it scores lowest of the four on DeepSWE v1.1 (64.6 against 74.2 for DeepSeek V4.1 Flash). That pattern fits the release story: search and visual work first, hard multi-step terminal work and repository level bug fixing still behind the frontier.

The jump from 2.0 is large on the agent tests Meituan chose to highlight. Terminal Bench v4.0 rises from 3.0 to 38.5. DeepSWE v1.1 rises from 9.7 to 64.6. Program Bench almost resolved rises from 2.5 to 25.5. AutomationBench rises from 15.8 to 51.8. Those are the same harness families, measured by the same vendor, so the direction is meaningful even if the absolute scores are not independently checked. A model that scored near zero on these tasks in June now scores in a range where it can be compared with other models. That is the honest version of the upgrade.

Two conditions change how you should use the numbers. First, all LongCat 2.5 scores were measured with thinking on, temperature 1.0, and top_p 0.95. Turn thinking off and the scores do not apply. Second, harnesses differ by test. Terminal Bench runs in Claude Code with a 512,000 token limit. DeepSWE runs in mini-swe-agent with a 1 million token limit. BrowseComp runs in Meituan's own lightweight harness with only search and browse tools, capped at 1,024 rounds. Scores from different harnesses are not directly comparable, even inside one table.

What the benchmarks do not prove

No independent reproduction was published as of November 3, 2026. Artificial Analysis, LMArena, and other independent leaderboards had not published a LongCat 2.5 score in the coverage checked for this page. Until one does, treat every score above as a vendor claim with a method note, not as a measured fact about your workload.

The one independent test found in coverage is narrow and useful. TokRace connected LongCat 2.5 to Claude Code on November 2 and ran 15 creative tasks: six 3D scenes, four videos, four static SVGs, and one animation, plus text and code checks. Its verdict was mixed. Clearly specified creative tasks already worked. Complex projects were unreliable and incomplete. Structural and explanatory details were often wrong: a gear face broken, a scale off by a factor of 1,000, a video that never arrived after timeouts. The strongest static result was a Chinese event poster with all six required strings in place. Thinking on generally gave tidier layouts but needed more requests. That is one test, on creative tasks, two days after release. It does not settle coding quality. It does warn against routing a long 3D or video job to this model without checking the output at each step.

One dispute is still open. AI Affairs reported that the 1.6 trillion total and 48 billion active figures are disputed, because a report before the full release said the figures attached to the preview came from the LongCat 2.0 model card. The official November 1 blog post states the same figures for 2.5 directly. This page uses the official post as the source and notes the dispute, because the preview history gives it weight. If Meituan publishes a technical report with architecture details, that report settles it. Until then, cite the figures as reported by Meituan, not as independently confirmed.

Finally, do not confuse the preview with the release when you read older coverage. Pages written in September and early October describe a preview with no benchmark table. Some still say the model is expected to follow 2.0 into open weights. The November 1 post does not promise weights or a license. Pages that present open weights as a fact for 2.5 are reading the 2.0 precedent as a guarantee. It is not one.

Pricing in full

The discounted price is the only price Meituan publishes on the API platform as a current rate. The table below shows it next to a reference price you can use for planning. The reference price is not Meituan's list price. Meituan has not published one.

Why $0.75, $0.015, and $2.95 appear as the reference set: those are the standard rates listed for LongCat 2.0 in third party listings, and one preview guide worked backwards from a 59 percent discount shown on a gateway page to about $0.73 input. A preview guide that checked the official price table in October states plainly that Meituan does not publish a list price. Use the reference set to test whether your project still works at roughly 2.5 times today's price. Do not quote it as Meituan's future price.

Billing details come from secondary checks of the platform, not from the blog post, so treat them as reported facts and confirm them on your own account. APIYI's November 2 guide reports that only requests returning HTTP 200 are billed, that failed requests such as 401, 403, 429, and 500 are not billed, and that prepaid balance does not expire, with promotional credits used first. The API docs confirm the rate limit behavior: an over-limit request returns HTTP 429 with error code rate_limit_exceeded and retry_after of 60, and the docs recommend exponential backoff. No RPM or TPM quota is published in the docs. If you need a fixed concurrency quota for a production service, ask before you commit, because the docs do not give you one.

For comparison at the discounted price, the same 800,000 input and 200,000 output token batch costs $0.48 on LongCat 2.5. APIYI's November 2 comparison prices the same batch at $0.62 on DeepSeek V4 Flash, about $2.00 on GLM-5.3, $5.40 on Kimi K3, $7.20 on GPT-5.6 Sol or Claude Opus 5.5, and $9.00 on Claude Opus 5. Those comparison prices are APIYI's, checked November 2. They will move. The useful point is the gap: at the discount, LongCat 2.5 is priced for high volume agent traffic, not for occasional chat. At the reference price, that gap mostly closes against DeepSeek V4 Flash, so the discount is doing real work in the comparison.

Budget warning: the discount has no published end date. A project that only works at $0.30 input and $1.20 output is a project that depends on a price Meituan can change. Run the calculator above at the reference price before you sign a customer up to a fixed fee.

How to connect it

The LongCat API docs list two endpoints. Use the docs domain, api.longcat.chat. Some third party guides published before the release list api.longcat.ai. That domain does not match the current docs, and the docs are the source to trust for endpoints.

OpenAI compatible, Python

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_LONGCAT_API_KEY",
    base_url="https://api.longcat.chat/openai",
)
response = client.chat.completions.create(
    model="LongCat-2.5",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1000,
)
print(response.choices[0].message.content)

Anthropic compatible, curl

curl -X POST https://api.longcat.chat/anthropic/v1/messages \
  -H "Authorization: Bearer YOUR_LONGCAT_API_KEY" \
  -H "Content-Type: application/json" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "LongCat-2.5",
    "max_tokens": 1000,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Claude Code

Point Claude Code at the Anthropic compatible endpoint. Set the auth token to your LongCat API key, the base URL to https://api.longcat.chat/anthropic, and the model to LongCat-2.5. Set the small, default, and Opus model variables to the same model, or Claude Code may still request its default models for subtasks. The docs also list integration guides for Codex, OpenCode, Kilo Code, Cline, and more than a dozen other tools. Check the exact field names in the guide for your tool before you paste settings in.

Keep your API key out of shared files and chat logs. The docs warn that a leaked key can be used to spend your quota. If a key leaks, create a new one on the API Keys page and delete the old one.

One practical limit to plan around: the context window is 1 million tokens, but the benchmark harnesses used 512,000 or 256,000 token limits for several tests, and the maximum output is 128,000 tokens. Some third party pages list 256,000 or 262,144 as the maximum output for the preview. The current API docs say 128,000. Build to 128,000, and check the docs again if a long output is cut short.

Preview to release: what changed, and what did not

June 30, 2026: LongCat 2.0 launches as a 1.6 trillion parameter coding agent model, trained on domestic accelerators. Weights ship MIT licensed on July 5.

September 25, 2026: LongCat 2.5 Preview reaches the LongCat API. It adds image understanding to the same 1.6 trillion scale and 1 million token context. Meituan publishes no benchmark table. The discounted price is already $0.30 input, $0.006 cached input, and $1.20 output. Existing verified users receive a token grant, reported as 5 million tokens.

November 1, 2026: the full release post, Introducing LongCat-2.5, publishes the training story, the visual tokenization design, the W4A4 inference work, eight end to end demos, and the 31 row benchmark table. GitHub and Hugging Face buttons say Coming soon.

November 2, 2026: news and guide coverage follows, plus the first independent test from TokRace. The price does not change. The weights do not appear.

What did not change matters as much as what did. The parameter scale, the context window, and the discounted price are the same as the preview. The release adds evidence and positioning, not a new size class or a new price. If you tried the preview in September, the November release gives you numbers to judge it by. It does not give you weights, a license, or a published rate limit.

Frequently asked questions

Sources

• LongCat official blog, Introducing LongCat-2.5 (November 1, 2026): https://longcat.ai/blog/longcat-2.5 (parameters, multimodality, training, W4A4 figures, 31 row benchmark table, weights marked Coming soon).
• LongCat API docs, Quick Start: https://longcat.ai/platform/docs (endpoints on api.longcat.chat, model name, 1M context and 128K maximum output, 429 handling).
• APIYI, LongCat 2.5 guide (November 2, 2026): https://help.apiyi.com/en/longcat-2-5-native-multimodal-agent-model-guide-en.html (discounted pricing, 1M token bill example, billing notes, Claude Code and Cline setup; its endpoint domain differs from the docs and is not used here).
• TokRace, LongCat 2.5 review (November 2, 2026): https://www.tokrace.com/en/guides/longcat-2-5-release (independent 15 task creative test, thinking on versus off, maximum output discrepancy in the docs).
• AI Affairs, Meituan releases LongCat-2.5 (November 2, 2026): https://www.aiaffairs.com/technology/meituan-releases-longcat-2-5-long-running-coding-tasks/ (release date, parameter dispute, preview history).
• Agmazon, LongCat-2.5-Preview explained: https://agmazon.com/blog/articles/technology/202610/longcat-2-5-preview-complete-guide-en.html (discount is limited time, no published list price, reference price caution, agent day cost example).
• ExplainX, LongCat 2.5 explainer: https://www.explainx.ai/blog/longcat-2-5-1-6-trillion-agent-model-2026 (preview history and list rates as reported in September; used only for that history, because its main text predates the November 1 benchmark table).

This page is an independent explainer, not affiliated with Meituan or LongCat. Figures that come from Meituan are labeled as reported by Meituan in the text. Prices were checked on November 3, 2026, and can change. The calculator runs in your browser. Nothing you type is sent to a server by this page.

More model guides

• Mistral Large 4, explained
• Gemini 4 Carbon, explained
• Kimi K3.1 Pro, explained
• DeepSeek V4.1 Flash 1017, explained