
Qwen4-Max vs LongCat 2.5: Two Flagships Nobody Has Scored, Twelve Times Apart on Price
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 96 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 130 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 152 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 119 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
Here is a comparison with an unusual property: neither side has a number. Qwen4-Max, generally available from Alibaba since 16 November 2026, publishes a rate card, a context window and a feature list, and no benchmark of any kind. LongCat 2.5, which Meituan published on 2 November 2026, ships with a 31-row vendor evaluation table and still has no independent score — there is no LongCat 2.5 entry on the Artificial Analysis leaderboard, only its predecessor LongCat 2.0, which carries no index value either. So when you put Qwen4-Max and LongCat 2.5 next to each other, the thing that actually separates them is not capability. It is $1.20 against $4.50 on output tokens, and it is whether you believe a vendor's own table when it is the only table there is.
That makes this page a different kind of comparison from the usual head-to-head. It is not asking which model is better, because nobody outside the two labs can answer that. It is asking which one you can afford to find out about, and what each vendor's documentation does and does not commit to.

What Meituan actually shipped on 2 November
LongCat 2.5 is a 1.6-trillion-parameter sparse mixture-of-experts model activating approximately 48 billion parameters per token, which makes it an iteration of the same architecture as LongCat 2.0 rather than a new design. It is natively multimodal in a specific and unusual sense: visual signals are encoded as multi-level discrete tokens through residual vector quantisation over features from a semantic-and-aligned encoder, so images and text enter the backbone as a single token vocabulary rather than through a bolt-on adapter. The vendor's stated use cases follow from that — building a page from a screenshot, generating a 3D scene from a description, producing a poster.
The serving contract is concrete. Context is 1,048,576 tokens in and 131,072 out. Two endpoints sit at api.longcat.chat: an OpenAI-compatible /openai/v1/chat/completions and an Anthropic-compatible /anthropic/v1/messages. Claude Code, Codex, Hermes, OpenCode and Kilo Code are named as supported harnesses with a first-party guide for each. Weights are not available; the blog's GitHub and Hugging Face buttons both read "Coming soon" with no date. LongCat 2.0 shipped under MIT, so the licence question for 2.5 is genuinely open rather than answered by precedent.

Alibaba's Qwen4-Max, by contrast, is closed in a different way: its documentation marks fine-tuning and model deployment as unsupported outright, which is a firmer statement than "coming soon". Text, image and video in; text out; a 1,000,000-token window; 131,072 tokens of output outside thinking mode; thinking on by default with a 131,072-token budget. Six regions — Singapore, Beijing, Hong Kong, Frankfurt, Virginia and Tokyo — against LongCat's single first-party platform.
The price gap, and what it buys
• Input — Qwen4-Max $1.50 per 1M in Singapore and $1.413 elsewhere, versus LongCat 2.5 $0.30
• Output — Qwen4-Max $4.50 Singapore and $4.24 elsewhere, versus LongCat 2.5 $1.20
• Cached input — Qwen4-Max $0.075 read after a $1.875 five-minute write, versus LongCat 2.5 $0.006 read
• Context — Qwen4-Max 1,000,000 tokens versus LongCat 2.5 1,048,576 tokens
• Max output — Qwen4-Max 131,072 tokens non-thinking, 262,144 shared with thinking, versus LongCat 2.5 131,072 tokens
• Modality — Qwen4-Max text, image and video input versus LongCat 2.5 text and image input, with native visual tokenisation
• Weights — neither publishes any, but LongCat 2.5 says "coming soon" while Qwen4-Max says unsupported
• Regions — Qwen4-Max six versus LongCat 2.5 one platform
• Independent score — neither has one
On output tokens that is a twelve-fold difference in Singapore and a 3.5-fold difference on input. Run the numbers on a real pipeline: a million calls a month at 2,000 input and 500 output tokens costs $5,250 on Qwen4-Max Singapore and $1,200 on LongCat 2.5. At ten times that volume the difference is $52,500 a month against $12,000, which is a headcount, not a rounding error.

What the extra money buys is not, on the published record, capability. It buys regional deployment — Frankfurt, Virginia and Tokyo matter enormously to a European or American company with a data-residency obligation, and Meituan's single platform does not offer them. It buys a video input path that LongCat does not document. And it buys the operational certainty of a vendor that has been selling frontier models to enterprises for years, against a platform whose own pricing page describes its figures as limited-time and closes by saying "prices are subject to change".
That last point deserves weight. LongCat's rate card tells you it is not the authoritative source for its own numbers. Qwen4-Max's rate card, published this morning, does not.
Reading a vendor table when it is the only table
Meituan's release page carries 31 evaluation rows, and the first interesting thing about them is the comparison set: GPT-5.5, GPT-5.6 Sol, Claude-Opus-4.8 and Claude-Opus-5. OpenAI's GPT-6.1 Sol, generally available since 29 September, is not in the table. Neither is Claude Opus 5.5, the model the opus alias in Claude Code resolves to. The opponents were chosen.
The second interesting thing is the summary sentence the announcement uses, that LongCat 2.5's scores sit "between Claude-Opus-4.8 and Claude-Opus-5". On the newest harness in the same table it sits 15 points below Opus 5. Both statements come from the same page. That is not a knock on the model — a 1.6T sparse model trained end to end on domestic Chinese accelerators and served at $1.20 output is a genuine achievement — but it is a reason to read the table as a chosen subset rather than a scorecard.
Qwen4-Max's problem is the inverse. There is no table at all. Alibaba's documentation specifies the mechanics with unusual precision — default temperature 1.0 with thinking and 0.7 without, top-p 0.95 and 0.8, top-k 20, a 131,072-token thinking budget, separate input ceilings of 991,808 tokens in non-thinking mode and 983,616 in thinking mode, six regions with their deployment scopes — and says nothing whatsoever about quality. No Qwen4 entry exists on the Artificial Analysis model index, and the per-model URL returns a 404.
So the choice between a table you have to discount and no table at all is closer than it looks. Both models require your own evaluation before production. The difference is that Meituan has given you a starting hypothesis and Alibaba has given you a specification.
The workload split that actually decides this
Three patterns, and the answer differs for each.
• High-volume text generation — LongCat 2.5 wins on arithmetic so decisively that the missing benchmark barely matters. At 3.5× to 12× cheaper, it can lose on quality by a noticeable margin and still win on cost per finished task. Run a hundred of your own prompts and check.
• Visual work from screenshots and screenshots-adjacent inputs — LongCat's native visual tokenisation is the reason the model exists, and it is the one claim in its release that is architectural rather than promotional. Qwen4-Max accepts video as well, so if your input is video rather than stills, it is the only one of the two that will take the request.
• Regulated or latency-sensitive deployment — Qwen4-Max's six regions against one platform is not a pricing consideration, it is an eligibility one. If your data has to stay in Frankfurt or Virginia, the twelve-fold price gap is irrelevant because the cheaper option cannot serve you at all.
Keeping both options open
OrcaRouter does not currently route LongCat 2.5 or Qwen4-Max; both are reachable through their vendors' own APIs, with the rate cards above. The models this comparison should be run against on our side are the ones already in the catalogue on the same key — GLM-5.3 at $1.40 and $4.40 per million, Qwen3.8-Max at $2.00 and $6.00, Kimi K3.1 Pro at $2.00 and $10.00 — because a three-way test of an unscored model against a priced and measured one is the cheapest way to find out whether the new model is worth its list price at all.
The routing layer is what makes that test cheap. A policy can send a sample of live traffic to LongCat 2.5 and to a known model in parallel, compare the outputs with whatever judge you trust, and only then decide whether to move the volume. Automatic failover matters more than usual with two models this new: neither has a track record long enough to justify a single-provider dependency, and a fallback chain costs nothing when the primary path is healthy.
The honest verdict
If your work is text and images at volume, LongCat 2.5 is the more interesting model on this page, and its twelve-fold output-price advantage is the only hard number either vendor has given you. If your work needs video input, regional deployment or a vendor with an enterprise track record, Qwen4-Max is the one that satisfies the requirement, and the price is what that satisfaction costs.
What neither model gives you is a reason to skip the evaluation. Two unscored flagships, published two weeks apart, priced a factor of ten apart, is exactly the situation in which an afternoon of your own traffic is worth more than every comparison table on the internet — including this one. The one thing to watch for on either side is the same: a model card rather than a changelog entry. Parameters, context, price and an API identifier in a single document is the signal that a model is finished. LongCat 2.5's weights button still says coming soon, and Qwen4-Max's page still has a blank where the benchmark table would go.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
