
Qwen4-Max Is Already Listed Outside Alibaba's API — at Alibaba's Own Global 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 · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 125 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 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 · 44 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 · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 134 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 · 128 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 · 186 tok/s
Qwen4-Max went generally available on 16 November 2026, and by the next morning it was listed somewhere other than Alibaba's own API. The listing quoted $1.413 per million input tokens, $4.24 per million output and $0.071 for a cache read. Those are not a reseller's numbers. They are Alibaba's Global-region card for Qwen4-Max, to three decimal places — the exact rates the vendor publishes for Beijing, Hong Kong, Frankfurt, US Virginia and Tokyo, which sit 5.8% below the Singapore card of $1.50 and $4.50. A model appearing on a second platform within a day of launch is not remarkable in 2026. A second platform that appears at the vendor's own list price tells you something specific about what is actually being served, and it is not a second deployment of a 2.4-trillion-parameter model.
What a same-price listing is evidence of
The read is narrow and it is worth stating as an inference rather than a fact, because nobody has published a capacity statement. Qwen4-Max is a 2.4T-parameter sparse mixture-of-experts with 71B activated per token plus 328B of N-gram embedding parameters, built on a new Qwen4 architecture with Gated DeltaNet and a sparse attention indexer. Standing that up on independent hardware takes procurement, kernels and a week of tuning at minimum. Matching the vendor's price to the thousandth of a dollar within a day of launch is what a pass-through looks like, and the cached-token rate is the tell: $0.071 is an odd number to arrive at independently and an obvious one to copy.
• Input — $1.413 per million on the third-party card, $1.413 on Alibaba's Global card, $1.50 in Singapore.
• Output — $4.24 on both Global-scope cards; $4.50 in Singapore.
• Cache read — $0.071 on the third-party card and on Alibaba's Global card; $0.075 in Singapore.
• What is not copied — the limits. The third-party page advertises a 977K-token context and a 128K maximum output. Alibaba's own card gives a 991,808-token input ceiling and a 262,144-token shared budget for thinking plus answer, of which 131,072 is available when thinking is off. A listing published this fast copies a price more reliably than it copies a ceiling.
That last bullet is the practical one. If you are sizing a long-context job — a 400K-token repository, a multi-hour agent trace — a platform that caps output at 128K will truncate work that the model itself would have finished. Price parity is not spec parity, and on a model this new there is no second source to check a limit against except the vendor's own model page.
The first independent benchmark exists. The first independent speed measurement does not.
This is the part of the story that matters more than the listing. Artificial Analysis added Qwen4-Max to its leaderboard with a full Intelligence Index row — 42.9, rounded to 43 on the public page, at $3.27 of measured cost per index task, having generated 127 million output tokens during the run, which the tracker ranks 65th of 80 on its verbosity scale. That is a real independent evaluation of a model that is two days old, and it is the number to quote when someone hands you a vendor chart.
What the same page does not have is a speed row. No output tokens per second, no time to first token, no time to first answer token. Every other flagship-tier model on that board carries those three figures; Qwen4-Max carries a null. That is not a criticism of the tracker — a throughput measurement needs repeated calls against a stable endpoint, and on 17 November there was no stable endpoint to call repeatedly. It is a warning about the reviews you will read this week: any article that quotes a tokens-per-second figure for Qwen4-Max in the next few days is either measuring a single request once or inventing it. The vendor has not published one either.

What an agent run actually costs, at these rates
Token prices are the wrong unit for a long-horizon model, so here is the arithmetic on a shape that Qwen4-Max is explicitly sold for — a coding agent that carries a 200,000-token context through 40 turns and writes 30,000 tokens of reasoning and patch per turn. All figures are list price, not a measured bill.
• Cached context, per turn — 200,000 tokens at $0.075 per million is $0.015. Across 40 turns, $0.60.
• Output, per turn — 30,000 tokens at $4.50 per million is $0.135. Across 40 turns, $5.40.
• The ratio — output costs nine times what the cached context costs, per turn, even after a 95% cache discount. The cache discount is not the lever people assume it is once output is long.
• Fresh input — every turn that misses the cache pays $1.50 per million instead of $0.075, a 20× jump. Four million tokens of cache misses across the run adds $6.00, more than the entire output bill.
Which inverts the usual advice. On Qwen4-Max the thing to instrument is cache hit rate, not output length: a 40-turn loop that keeps its context warm lands near $6.00, and the same loop that re-sends its context cold lands near $12.00. The vendor's own headline — average cost per task on its Work360 benchmark falling from $4.43 to $1.67 — is a claim about its benchmark's token mix, not a promise about yours. It is also a vendor number on a vendor benchmark, and it is labelled as such here because nothing independent reproduces it yet.

Trying it without re-platforming
OrcaRouter does not route Qwen4-Max. That is worth saying plainly rather than burying: the catalogue carries qwen/qwen3.8-max, anthropic/claude-fable-5.2 and openai/gpt-6.1-sol, and a request for a qwen4 identifier returns not-found today. If you want the new Qwen flagship on our key, you are waiting like everyone else.
What this situation does argue for is not switching your whole stack to a two-day-old model to test it. A preview-to-GA transition that also cut the shared output budget from 393,216 tokens to 262,144 and moved billing from credits to per-token metering is exactly the kind of change that breaks a production path in week two. That is the case for keeping the experiment behind a routed endpoint rather than a hardcoded one — new model on one route, the incumbent on the fallback, one bill either way. When Qwen4-Max lands in the catalogue it will be listed at the vendor's price with 0% markup, which means the same pass-through that makes a third-party card identical to Alibaba's is the thing that makes a vendor price cut visible on our side the day it happens.

The fine print has a date on it
Two identifiers are on a clock. qwen4-max-preview and qwen4.0-max-preview stop working at 24:00 on 25 November 2026, UTC+8 — nine days after general availability, and the model identifier that survives is plain qwen4-max. Any config, SDK default, notebook or eval harness still pointing at either preview string will fail with a model-not-found rather than a deprecation warning, which is the least helpful way to find out.
So the story is smaller and more useful than "it is available everywhere now." A model shipped, a third-party card appeared within a day at the vendor's own Global rates, and the limits attached to that card are not the vendor's limits. The independent intelligence score arrived within a day; the independent speed score has not arrived at all. Read the price as a pass-through, read the ceiling off the vendor page, and put a calendar reminder on 25 November for the preview identifiers — that is the whole of what 16 November actually told you.
