China · LLM · GA · Released Nov 2, 2026

LongCat 2.5: The Topic Page

Meituan LongCat's latest-generation model, GA on November 2, 2026: a 1.6T-parameter mixture-of-experts with about 48B active per token, native multimodality (text, image and video input), trained for long-horizon agent work, with a 1M-token context. Its limited-time price of $0.30/$1.20 matches DeepSeek V4.1 Flash exactly, and Meituan self-reports 88.4 on Terminal-Bench 2.1 — but as of November 13 no independent lab has scored it, the discount has neither an end date nor a list price, and the weights quietly went MIT open-source on November 4, a change most early coverage hasn't caught up with. This page collects the official specs, the ways to try it, pricing, and the first hands-on reviews in English and Chinese — with unverified claims clearly marked.

1.6T / 48BTotal params / active per token

88.4Terminal-Bench 2.1 (vendor-reported)

$0.30 / $1.20Promo price per 1M tokens (in / out)

1M tokensContext (API figure)

MIT, open weightsWeights out Nov 4, 3.3 TB full release

No independent scoreNot yet rated by AA (as of Nov 13)

01Key facts

What is confirmed: a 1.6T-parameter native multimodal agent model, priced head-to-head against DeepSeek V4.1 Flash, with MIT-licensed weights already out — but every eye-catching score is vendor-reported, with no independent verification at all.

Released: 2026-11-02 GA (live on the LongCat API Platform; tech blog announced it Nov 1, preview Sep 25). This site uses Nov 2 as the release date: the first day it was actually usable.
Developer: Meituan LongCat team, the large-model team of Meituan, China
Architecture: Sparse MoE; 1.6T total parameters, about 48B active per token (weights repo config.json: 38 layers, 768 routed experts, top-12 routing, MLA attention, 163,840-token vocabulary)
Context: 1M tokens (API figure 1,048,576, input + output combined); config base 524,288, extended via YaRN; max output 128K (documentation disagrees — see section 09)
Input / output: Input: text, image, video. Output: text. Reasoning returns via a reasoning_content field; temperature is fixed (1.0 with thinking on, 0.7 with it off — other values are silently overridden)
New capabilities: Native multimodal vision (multi-level discrete tokens, no separate vision encoder); video understanding added at GA; long-horizon agent training (F2P/P2P reinforcement learning, MOPD multi-teacher distillation)
Price: Limited-time discount: $0.30 in / $0.006 cached / $1.20 out per 1M tokens (¥2 / ¥0.04 / ¥8); no end date, no list price — Meituan says "prices may change"
Benchmarks: 31 vendor-reported tests: Terminal-Bench 2.1 88.4, DeepSWE 64.6, BrowseComp 91.9 (table-best), 3DCodeBench 55.4 (table-best); independent score: none (Artificial Analysis has not rated it, as of Nov 13)
Weights: Released (2026-11-04, MIT license; full release 3,317.3 GB, W4A4 quantized build 1,244.3 GB; Hugging Face, plus a ModelScope mirror)
Access: longcat.ai/chat, LongCat API Platform (model ID LongCat-2.5), Vercel AI Gateway, AnyRouter, OpenCode (free Zen tier), Nous Portal; not on OpenRouter (checked Nov 13)
Predecessor: LongCat 2.0 (GA 2026-06-30, MIT weights); 2.5 shares 2.0's base model, adding vision and agent training
Knowledge cutoff: Not published by Meituan

02Timeline

Four months from 2.0 to 2.5: preview, announcement, GA and open weights inside seven weeks — the weights landed without even a formal announcement.

2026-04-20
LongCat 2.0 preview launches: a text-only agent model with 1M context.
2026-06-30
2.0 goes GA with MIT weights (Hugging Face repo created Jul 5); list price at the time $0.75 / $0.015 / $2.95.
2026-09-25
LongCat 2.5 preview launches: text-only, vision not yet enabled; a community self-test of the preview scored about 48.7% on Terminal-Bench 2.1 (single developer, different harness — reference only).
2026-11-01
Tech blog announces LongCat 2.5: 1.6T parameters, native multimodal architecture and 31 official benchmarks; weights marked "coming soon."
2026-11-02
GA on the LongCat API Platform: video understanding added at GA; the $0.30/$1.20 limited-time price takes effect the same day. Meituan's tech team blog publishes the training method (domestic accelerators, W4A4, long-horizon RL) the same day.
2026-11-04
Weights quietly land on Hugging Face: full + W4A4 repos, MIT license, no formal announcement; Nous Portal launches a free endpoint the same day.
2026-11-09
OpenCode Zen launches a free tier, longcat-2.5-free; the paid version ships with OpenCode Go.
2026-11-13
Artificial Analysis still hasn't rated it (ArtificialWatch's report that day: "not rated yet"); this topic page goes live.

03Official specs

As published by Meituan (tech blog, API Platform documentation and the Hugging Face weights repo's config.json, checked Nov 13, 2026). Note: max output and context each have two figures — see the text and section 09.

Status: Released (GA 2026-11-02); weights released (2026-11-04)
Model ID: LongCat-2.5
Architecture: Sparse MoE; 1.6T total parameters, about 48B active; 38 layers, 768 routed experts, top-12 routing, MLA attention, 163,840-token vocabulary
Context window: 1M tokens (API figure 1,048,576, input + output combined; config base 524,288, extended via YaRN rope scaling)
Max output: 128K tokens (Quick Start, FAQ and Meituan's Claude Code config; the Model Parameters page says max_tokens can reach 262,144 — the docs disagree, so plan on 128K)
Input: Text, image (PNG/JPG/JPEG/WebP/GIF, ≤10MB, ≤50 per request), video (≤100MB, <3600s, ≤10 per request, public URL only; no SVG/HEIC; aspect ratio >200:1 errors)
Output: Text
Reasoning: Returned via a reasoning_content field; temperature fixed (1.0 with thinking on, 0.7 with it off)
API protocols: OpenAI Chat Completions, OpenAI Responses (for Codex, effort none to xhigh), Anthropic Messages; endpoint api.longcat.chat
Training: Trained on domestic accelerators (<80GB VRAM); W4A4 quantized inference; long-horizon RL (F2P/P2P, anti-reward-hacking); MOPD multi-teacher distillation
Benchmarks: 31 vendor-reported tests (see section 06C); no independent score
Weights: MIT; full release 3,317.3 GB (210 shards), W4A4 1,244.3 GB (208 shards)
License: MIT
API price: $0.30 in / $0.006 cached / $1.20 out (limited-time discount, no end date)

04How to try it: six routes

The weights are open, but a 3.3 TB download and NPU-oriented deployment docs mean most people will still use a hosted API. Fastest free starts: OpenCode's free tier and the Nous Portal endpoint.

1LongCat API Platform (official main route)

Meituan's official API platform, model ID LongCat-2.5. OpenAI- and Anthropic-compatible, endpoint api.longcat.chat; prepaid balance never expires, only HTTP 200 requests are billed. Email signup, no identity check outside mainland China; the USD site takes credit cards, debit cards and Google Pay. Thirteen official tool setup guides (Claude Code, Codex, OpenCode, Cline, Kilo Code, Hermes, OpenClaw and more).

2longcat.ai/chat (official chat app)

The official chat entry point — sign in and use it; replies default to Mandarin. Some community feedback notes the login redirects to another domain, which feels disjointed.

3Gateways: Vercel AI Gateway, AnyRouter

2.5 is live on Vercel AI Gateway and AnyRouter, useful for developers who already bill through a gateway. Note: not on OpenRouter (checked via the OpenRouter API Nov 13 — only meituan/longcat-2.0 is listed); APIYI has announced plans to add it but had not as of Nov 3.

4OpenCode Zen free tier

Since Nov 9, OpenCode Zen offers a free tier, longcat-2.5-free; the paid version ships with OpenCode Go. A zero-cost trial for coding-agent work.

5Nous Portal free endpoint

A free endpoint went live Nov 4 — another zero-cost trial route.

6Self-hosting (weights are open)

Both Hugging Face repos are downloadable: the full release at 3,317.3 GB (210 shards) and W4A4 at 1,244.3 GB (208 shards), MIT license, with a ModelScope mirror. The README only documents NPU deployment via Meituan's SGLang-FluentLLM fork; vLLM support has no official statement (unverified), and Meituan publishes no VRAM guidance — don't estimate GPU counts from the 3.3 TB size.

Official API: api.longcat.chat, model ID LongCat-2.5 (OpenAI + Anthropic compatible)
Free tiers:  OpenCode Zen (longcat-2.5-free), Nous Portal free endpoint
Gateways:   Vercel AI Gateway, AnyRouter (not on OpenRouter)
Weights:    Hugging Face meituan-longcat/LongCat-2.5 (MIT, full 3.3 TB / W4A4 1.24 TB)

Before you start: community feedback reports looping on long tasks, with performance dropping noticeably past about 200K context (Reddit preview feedback, a small sample — treat as anecdotal); Meituan publishes no data policy — the FAQ says nothing about whether prompts are retained or used for training, so assess sensitive use cases (such as customer support) yourself. Note also that 2.0 and 2.5-Preview have disappeared from the official model list, with no dated retirement notice.

05Pricing

The promo price matches DeepSeek V4.1 Flash's peak price to the cent — a deliberate head-to-head. But the discount has no end date and no list price, so long-term budgets can't be locked in at this rate. Prices checked Nov 13, 2026; the checkout page is authoritative.

Item (per 1M tokens)	Promo price	RMB price
Input (uncached)	$0.30	¥2
Cached input	$0.006	¥0.04
Output	$1.20	¥8
Volume discounts (verified companies, monthly spend $75+ / $300+ / $1,200+ / $6,000+)	10% / 20% / 30% / 40% off	—
Token Packs (30-day quota, flash drops at 10:00 / 16:00 / 21:00 / 23:00 Beijing)	Sizes and prices behind login	—

The cache price is the real weapon

Cached input at $0.006 is just 2% of the uncached rate (1/50) — a big deal for agent tasks that repeatedly read the same repo or long document. Meituan's worked example: 600K uncached input + 400K cached input + 500K output = $0.7824. Billing is friendly too: only HTTP 200 requests are charged, prepaid balance never expires, and there is no context-length surcharge and no image/video surcharge.

Same-price comparison

DeepSeek V4.1 Flash's peak price of $0.30/$1.20 is identical to LongCat's promo price, but DeepSeek halves its rates off-peak ($0.15/$0.60), making its monthly bill cheaper; same-tier flagships cost several times more: Kimi K3.1 Pro $2/$10 and Qwen3.8 Max $2/$6 (both official API prices, checked Nov 13).

Pricing bottom line: no end date, no list price — Meituan only says "prices may change," which is both an opportunity and a risk. If the price later moves to 2.0's old list of $0.75/$2.95, output would cost about 2.5× more (speculation — Meituan has not said so). Don't lock long-term budgets to the promo price; a community post claims 10M tokens of starter credit for new users (5M for existing) plus a ¥9.9 offer "coming soon" — unverified.

Official pages: Meituan official tech blog · Hugging Face weights repo · LongCat official site

06Hands-on reviews

Everything below is a first read or hands-on from Nov 1–13: small samples, mixed methods — treat each as a case study. Watch the date trap: anything written before Nov 4 says the weights are "coming soon," which is now outdated. Each entry carries a verdict badge; "Read the local article" opens this site's rewritten deep dive, with the original source at the end.

Chinese sources

Meituan tech team blog: how the long-horizon agent was trained

Verdict: the only detailed disclosure of the training method

Source: Meituan tech team blog, 2026-11-02 (training: long-horizon RL, MOPD distillation, W4A4 quantization)

Long-horizon reinforcement learning uses an F2P/P2P framework with anti-reward-hacking design; multi-teacher distillation (MOPD) compresses several teachers into one model; W4A4 quantized inference is about 43% faster on both prefill and decode, with time to first token down about 16%. Training ran entirely on domestic accelerators, each with under 80GB of VRAM.

Beijing News: a trillion parameters, trained on domestic accelerators

Verdict: Chinese press coverage of the training infrastructure

Source: Beijing News (translated by TokenHeart), 2026-11-02

The report focuses on the 1.6T parameter scale and the all-domestic-accelerator training run; it adds no benchmark methodology or pricing detail beyond the official blog.

English sources

eesel deep dive: a cheap vision coding agent, but same-price DeepSeek is stronger

Verdict: worth trying, but don't expect it to beat DeepSeek

Source: eesel, 2026-11-03 (page-by-page read of the official docs: vision limits, output cap, community feedback, data policy)

Vision input is 2.5's real increment, but the output cap is disputed — 128K vs 262,144 — and eesel advises planning on 128K; Reddit reports looping on long tasks and degradation past about 200K context; the FAQ says nothing about prompt retention or training use, a red flag for support use cases. Note: eesel wrote on Nov 3, so its "weights unreleased" conclusion is outdated since the Nov 4 MIT release.

Read the local article → (our own write-up)

apiyi benchmark guide: 25 tests, item by item

Verdict: the fullest reading of the official table, with pricing math and setup steps

Source: apiyi, 2026-11-03 (official benchmark table, pricing examples, setup guide)

Item-by-item coverage of the official benchmark table with a full breakdown of the $0.7824 worked example; a companion piece explains the long-horizon agent training. Also written before the weights release, so its "weights coming soon" wording is outdated.

Read the local article → (our own write-up)

ToolPilotPro three-way comparison: careful bill math, half-outdated conclusions

Verdict: the bill math is solid; use our corrected version for the openness verdict

Source: ToolPilotPro, updated 2026-11-04 (LongCat vs Kimi vs DeepSeek)

Monthly workload math (40M input, half cached, + 8M output): DeepSeek $7.86 off-peak / $15.72 peak, LongCat $15.72, Kimi $124. But the Nov 4 update still calls LongCat API-only with weights "coming soon" — the MIT weights have been on Hugging Face since Nov 4, and our comparison is corrected accordingly.

Read the local article → (our own write-up)

OrcaRouter preview analysis: a preview-era check

Verdict: preview data — reference only after GA

Source: OrcaRouter, data from 2026-10-13 (preview verification)

An early check of the preview's specs and pricing; vision, video input and the weights status have all changed since GA, so post-GA sources take precedence.

TokenDyno live speed test: about 51.8 tok/s

Verdict: mid-pack speed, tiny sample

Source: TokenDyno, 2026-11-13 (OpenCode Go, n=4)

LongCat 2.5 (Free) on OpenCode Go runs about 51.8 tok/s with a time to first token of about 2.8 seconds. Only four samples, and gateway and load affect results — reference only.

Community feedback: looping, the comparison set, and one self-test

Verdict: anecdotes, but they all point at the same gap — no independent reproduction

Source: Reddit, Hacker News, from 2026-11-02

Reddit preview feedback: loops on long tasks, good in narrow roles on small tasks, not suited to planning or creative writing; HN criticized the price and the comparison set (the headline chart omits Opus 5.5, GPT-6.1 Sol and Kimi K3.1 Pro); a single developer self-tested the preview at about 48.7% on Terminal-Bench 2.1 — a different harness and version from the official 88.4, so not directly comparable.

06BArticles: our in-depth coverage

The six articles below are our own rewritten and reorganized deep dives — restructured, with added conclusions and selection advice, and facts and figures checked against the original reviews. Open any one to read it on this site; the original source is linked at the end.

LongCat 2.5 Launches: 1.6T Multimodal Agent ModelEverything confirmed about the Nov 2 GA: what's actually new in native vision and video input, long-horizon RL and MOPD distillation, the relationship to 2.0, and the Nov 4 MIT weights reversal. LongCat 2.5 Benchmarks: 31 Official Tests ReviewedAll the official scores: 31 tests in three groups, item by item, the same-price DeepSeek comparison, the 2.0-to-2.5 jump, and the missing AA score plus the 20-point harness warning. LongCat 2.5 Pricing: $0.30/$1.20 Promo and the CatchThree bills worked out: why cached input at 2% of the uncached rate matters, the four volume-discount tiers, the $0.7824 worked example, and the price-rise risk of a discount with no end date and no list price. eesel on LongCat 2.5: DeepSeek Wins Most Coding RowsA deep read of eesel's page-by-page review: the real value of vision input, the 128K output-cap dispute, community looping reports, the missing data policy as a support-use red flag, and four-tier selection advice. LongCat 2.5 vs DeepSeek vs Kimi: Which One to PickA three-way comparison on price, benchmarks, licensing and availability — correcting ToolPilotPro's outdated conclusions after the Nov 4 weights release, with scenario-based selection advice. LongCat 2.5 Weights Out: MIT License, 3.3 TB DownloadThe weights situation, verified: Early-November coverage said "coming soon" — outdated. Both repos' creation times, sizes and download counts, what MIT means, and the NPU deployment and self-hosting reality.

06CBenchmarks: 31 official tests, no independent measurement (yet)

Meituan published 31 tests in three groups — code agent, general agent, visual — and the data is thorough; but all of it is vendor-run, and rival scores are half public reports, half Meituan's own runs. Read the conditions first, then the independent measurements — currently, there are none.

Test (official table, vendor-run)	2.5	2.0	Same-price rival: DeepSeek V4.1 Flash
Terminal-Bench 2.1	88.4	70.8	90.6
Terminal-Bench 4.0	38.5	3.0	31.2
DeepSWE v1.1	64.6	9.7	74.2
BrowseComp	91.9	—	91.5
3DCodeBench Text→3D	55.4	—	—
PaperBench	82.1	—	—
VideoMME	85.2	—	—
HLE (with tools)	63.1	—	—
JobBench	65.4	—	—
MCP Atlas	83.1	—	84.4
AutomationBench	51.8	15.8	54.8
NL2Repo	60.5	—	65.4
ProgramBench	25.5	2.5	20.3
Agents' Last Exam v1.1	38.5	—	—
ResearchRubrics	81.0	—	—

Independent measurement and third-party data

Measurement	Source	2.5 result	Notes
AA Intelligence Index	Artificial Analysis	Not rated (as of Nov 13)	The AA list has only 2.0 and Flash Lite
ArtificialWatch report	ArtificialWatch, Nov 13, 08:59 UTC	"not rated yet"	—
TokenDyno live speed test	TokenDyno (OpenCode Go, n=4)	About 51.8 tok/s, TTFT about 2.8s	—
OpenVibeEval ranking	OpenVibeEval leaderboard	2.5 Preview ranked No. 40	Preview, not the GA release
Community self-test, TB 2.1	Single developer, preview, different harness	About 48.7% pass@1	Not comparable to the official 88.4

Read the conditions first: all 31 official tests were run with thinking on, temperature 1.0 and top_p 0.95; Terminal-Bench used Claude Code v2.1.280 (max sequence 512K), DeepSWE used mini-swe-agent (1M context). Rival scores marked * come from public reports or leaderboards; unmarked ones are Meituan's own runs, labeled "indicative only" by Meituan itself. Warning: AA once measured 2.0's TB 2.1 at 50.2%, 20 points below Meituan's table figure of 70.8 — a different harness can move a score by an order of magnitude, so treat 2.5's 88.4 as a vendor figure until an independent reproduction exists. Note also that the headline chart omits Opus 5.5, GPT-6.1 Sol and Kimi K3.1 Pro, which the community has criticized.

Full coverage in our benchmarks article →

Official benchmark scores (percent; different tests, not directly comparable)

Meituan official LongCat 2.5 metrics dashboard: code, general-agent and visual benchmarks vs GPT-5.5, GPT-5.6-Sol, Claude-Opus-4.8 and Claude-Opus-5 — Meituan official metrics dashboard. Source: Meituan official tech blog

07Comparisons

vs LongCat 2.0 (predecessor)

Dimension	LongCat 2.5	LongCat 2.0
Released	GA 2026-11-02	GA 2026-06-30
Parameters	1.6T total / about 48B active	Same base model (2.5 shares 2.0's base)
Input	Text, image, video	Text only
Weights	MIT, released (Nov 4)	MIT, released
Terminal-Bench 2.1	88.4	70.8 (AA measured 50.2)
Terminal-Bench 4.0	38.5	3.0
DeepSWE v1.1	64.6	9.7
AutomationBench	51.8	15.8
ProgramBench	25.5	2.5
Price	$0.30 / $1.20 (promo)	$0.75 / $2.95 (old list price)

The 2.5 jump comes from vision and agent training, not scale — same base model, yet Terminal-Bench 4.0 went from 3.0 to 38.5 and DeepSWE from 9.7 to 64.6. And the price went down: the promo rate is about 40% of 2.0's old list price.

vs same-price and same-tier rivals (as of 2026-11-13)

Model	Released	AA Index	API price $/1M (in / out)	Weights
Meituan LongCat 2.5	Nov 2	Not rated	0.30 / 1.20 (promo)	MIT, released
DeepSeek V4.1 Flash	Sep 10	32.6	0.30 / 1.20 (peak); 0.15 / 0.60 (off-peak)	MIT, released
Kimi K3.1 Pro	Oct 23	46	2.00 / 10.00	Released Nov 4 (custom license)
Tencent Hy4	Nov 4	35	0.83 / 2.50 (50% off: 0.42 / 1.25)	Pending
Inkling 2 Small	Nov 5	31	0.30 / 1.00	Apache 2.0, released

An honest note: vendor-reported scores use different versions and harnesses and can't be compared directly — DeepSWE is 64.6 (LongCat), 77.0 (DeepSeek) and 75.5 (Kimi) across the three vendors' tables, and they aren't the same test version. LongCat's AA Index is "not rated," not a low score — don't read it as zero. On price, LongCat matches DeepSeek's peak rate and is close to Inkling; Kimi costs about 7–8× more.

Read the deep comparison: LongCat 2.5 vs DeepSeek & Kimi →

08What's missing (as of Nov 13)

An Artificial Analysis score (not rated as of Nov 13; no third-party benchmark reproduction exists)
A data policy: whether prompts are retained or used for training, unaddressed in the FAQ
The promo discount's end date and the post-discount list price
Token Pack sizes and prices (behind login, unpublished)
Rate limits (RPM/TPM, unpublished)
vLLM support and self-hosting VRAM guidance (the README only covers NPU + SGLang-FluentLLM)
Knowledge cutoff date (not published)
A dated retirement notice for 2.0 / 2.5-Preview (gone from the model list, no formal notice)
Independent long-context evaluation of the 1M window (config base 512K + YaRN, unverified in practice)
Reconciliation of the 128K vs 262,144 max-output figures in the docs

09Rumor check

Most LongCat 2.5 confusion comes from timing — coverage from Nov 1–3 was largely outdated by the Nov 4 weights release — plus a few misreported numbers, clarified item by item below.

False "440B active parameters"

Some launch-day community posts claimed 440B active — wrong. The official figure is about 48B active (config.json: 768 routed experts, top-12 routing); 440B has no official source.

Clarified "Weights coming soon — API only"

Every report from Nov 1–3 said this (including ToolPilotPro's comparison, updated Nov 4), but both repos have been on Hugging Face since Nov 4: the full release at 3,317.3 GB and W4A4 at 1,244.3 GB, MIT license. This site verified the creation times and licenses via the HF API.

False "$0.30/$1.20 is the permanent price"

The official pricing page labels it a limited-time discount, with no end date and no list price, adding "prices may change." 2.0's old list price was $0.75/$2.95 — if 2.5 moves there, output would cost about 2.5× more (speculation; Meituan has not said so).

Partly true "Terminal-Bench 2.1 is only 48.7%"

48.7% is a single developer's self-test of the preview on a different harness — not the GA release's score; the official 88.4 is likewise a vendor figure. The two numbers don't measure the same thing, and neither is a verdict on the GA release until an independent reproduction exists.

Partly true "1M context"

The API figure is indeed 1,048,576 (input + output combined), but the config base is 524,288, extended to 1M via YaRN. 2.0 had a history of "1M advertised, 256K in the config" (eesel), and 2.5's 1M has no independent long-context evaluation yet.

Clarified "Was it released Nov 1 or Nov 2?"

The tech blog announced it Nov 1; the API Platform went GA Nov 2. This site uses Nov 2 as the release date: the first day it was actually usable. The preview was Sep 25.

10FAQ

When was LongCat 2.5 released?

Announced on the tech blog Nov 1, 2026, and GA on the LongCat API Platform Nov 2; the preview shipped Sep 25. This site uses Nov 2 as the release date. See the timeline in section 02.

Where can I download the LongCat 2.5 weights?

From two repos under the meituan-longcat organization on Hugging Face: LongCat-2.5 (full release, 3,317.3 GB, 210 shards) and LongCat-2.5-W4A4 (1,244.3 GB, 208 shards), both MIT-licensed, published Nov 4, 2026; ModelScope also mirrors them. See our weights status article.

Is LongCat 2.5 open source?

Yes. The weights were released Nov 4, 2026 under the MIT license — commercial use, modification and redistribution are all allowed; 2.0 is MIT too. Note that much early-November coverage still says the weights are "coming soon" — that's outdated.

How much does LongCat 2.5 cost?

Promo pricing: $0.30 input / $0.006 cached input / $1.20 output per 1M tokens (¥2 / ¥0.04 / ¥8). The discount has no end date and no list price, and Meituan says "prices may change." Only HTTP 200 requests are billed; prepaid balance never expires. See our pricing guide.

How do I use LongCat 2.5, and where is it available?

The official chat app at longcat.ai/chat, the LongCat API Platform (model ID LongCat-2.5), Vercel AI Gateway and AnyRouter; free routes include OpenCode Zen (longcat-2.5-free) and the Nous Portal free endpoint. Note: it's not on OpenRouter. Self-hosting is possible with the MIT weights, but the README only documents NPU + SGLang-FluentLLM.

What is LongCat 2.5's AA Index score?

As of Nov 13, Artificial Analysis has not rated LongCat 2.5 — there is no independent score; the AA list has only 2.0 and Flash Lite. ArtificialWatch's Nov 13 report likewise says "not rated yet." This site will keep tracking it.

Can I trust the 88.4 on Terminal-Bench 2.1?

88.4 is Meituan's self-reported figure (Claude Code v2.1.280 harness, max sequence 512K) — treat it as a vendor figure until an independent reproduction exists. Warning: AA once measured 2.0's TB 2.1 at 50.2%, 20 points below Meituan's table figure of 70.8; a community self-test of the 2.5 preview scored about 48.7% on a different harness (reference only).

How does LongCat 2.5 differ from 2.0?

They share a base model; 2.5 adds native vision (image + video input) and long-horizon agent training: Terminal-Bench 4.0 rose from 3.0 to 38.5, DeepSWE from 9.7 to 64.6, AutomationBench from 15.8 to 51.8. The price went down instead: the 2.5 promo of $0.30/$1.20 is about 40% of 2.0's old list price of $0.75/$2.95.

How does LongCat 2.5 compare with DeepSeek V4.1 Flash?

The promo/peak prices are identical ($0.30/$1.20), but DeepSeek halves its rates off-peak, so its monthly bill is cheaper; on benchmarks DeepSeek leads most coding rows (DeepSWE 74.2 vs 64.6, TB 2.1 90.6 vs 88.4), while LongCat leads TB 4.0 (38.5 vs 31.2), BrowseComp (91.9 vs 91.5) and ProgramBench (25.5 vs 20.3). Both weights are MIT. See our three-way comparison.

Does LongCat 2.5 support vision and video?

Yes. Input can include images (PNG/JPG/JPEG/WebP/GIF, ≤10MB, ≤50 per request) and video (≤100MB, <3600s, ≤10 per request, public URL only); SVG/HEIC are not supported, and an aspect ratio >200:1 errors. Vision uses native multi-level discrete tokens with no separate vision encoder; output is still text-only. Video understanding arrived with the Nov 2 GA — the preview didn't have it.

Is max output 128K or 262K?

The official docs disagree: the Quick Start, FAQ and Meituan's Claude Code config say 131,072 (128K), while the Model Parameters page says max_tokens can reach 262,144. This site's advice, matching eesel's: plan on 128K; 262K is unverified.

How large is LongCat 2.5's context window?

The API figure is 1,048,576 tokens (input + output combined). The weights repo's config base is 524,288, extended to 1M via YaRN rope scaling; real-world long-context performance has no independent evaluation yet.

11Sources

Every fact on this page comes from the public sources below (checked Nov 13, 2026).

LongCat 2.5: The Topic Page

01Key facts

02Timeline

03Official specs

04How to try it: six routes

1LongCat API Platform (official main route)

2longcat.ai/chat (official chat app)

3Gateways: Vercel AI Gateway, AnyRouter

4OpenCode Zen free tier

5Nous Portal free endpoint

6Self-hosting (weights are open)

05Pricing

The cache price is the real weapon

Same-price comparison

06Hands-on reviews

Chinese sources

English sources

06BArticles: our in-depth coverage

06CBenchmarks: 31 official tests, no independent measurement (yet)

Independent measurement and third-party data

07Comparisons

vs LongCat 2.0 (predecessor)

vs same-price and same-tier rivals (as of 2026-11-13)

08What's missing (as of Nov 13)

09Rumor check

False "440B active parameters"

Clarified "Weights coming soon — API only"

False "$0.30/$1.20 is the permanent price"

Partly true "Terminal-Bench 2.1 is only 48.7%"

Partly true "1M context"

Clarified "Was it released Nov 1 or Nov 2?"

10FAQ

11Sources

Official

English coverage and analysis

12Related topics