
Qwen4-Max Was Listed Elsewhere Within a Day. The Thinking Budget Is the Part That Will Not Port
- 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 · 133 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 · 193 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 · 126 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 · 185 tok/s
Qwen4-Max went generally available on 16 November 2026, and within hours a second platform was advertising it from day zero — 1M-token context, native text, image and video input, pitched at coding, office work and long-horizon agent runs. That is the normal shape of a closed-model launch in 2026: Alibaba publishes no checkpoint for Qwen4-Max, so anyone outside its six deployment regions can only reach the model through a platform that rents it from Alibaba. What is not normal, or at least not discussed, is how much of the model's operating manual gets lost in that relay. The price card travels cleanly. The headline context number travels cleanly. The arithmetic that decides whether your request fits or returns a truncation does not travel at all.
What a day-zero listing actually tells you
Two things, neither of them about benchmarks. First, that the model is being sold into research and agent work rather than chat — the day-zero copy leads with long-horizon agent runs and lists video as an input modality, which is a description of the workload the platform intends to point it at, not a restatement of Alibaba's own headline. Second, that Qwen4-Max is rentable but not portable: there is no download, so every endpoint serving it is serving Alibaba's deployment, and every third-party page describing it is a summary of Alibaba's documentation. Summaries of spec tables are where integration bugs are born.
The rate card is the easy part and worth stating once. In Singapore, Alibaba lists Qwen4-Max at $1.50 per million input tokens and $4.50 per million output, with cached input at $0.075 and cache writes at $1.875 for the five-minute window or $3.00 for the hour. The Global deployment scopes — Beijing, Hong Kong, Frankfurt, Japan and US Virginia — run slightly lower at $1.413 in, $4.24 out and $0.071 for a cache read. It is a single pricing tier all the way to the context limit: no long-context surcharge, no batch discount to hunt for, because batch inference is not offered on this model at all.

Three numbers that do not travel with the model
"1M context" is the number on every listing. Alibaba's model page, last updated on launch day, is more specific, and the gap between the marketing number and the operative ones is where integrations break:
• Context length — 1,000,000 tokens, a ceiling on the whole request rather than a budget you can spend on output
• Maximum input, non-thinking — 991,808 tokens, because 8,192 of the window are held back for the reply
• Maximum input, thinking — 983,616 tokens, a further 8,192 held back for the reasoning block
• Maximum output, non-thinking — 131,072 tokens
• Combined thinking plus answer, thinking mode — 262,144 tokens, shared between the reasoning trace and the visible reply
• Default thinking budget — 131,072 tokens, with reasoning_effort defaulting to high when neither parameter is set
The fifth line is the one that bites. A team that sizes a job against a documented 262,144-token output ceiling and then runs it in thinking mode will see the visible answer stop at a length that has nothing to do with how much the model had left to say, because the reasoning trace has been drawing from the same pool. The 131,072 figure is the safe output number. The 262,144 figure is a shared budget, and the difference between those two readings is the difference between a working pipeline and a weekly incident.
There is a small consolation: this also explains a discrepancy you will see in third-party spec listings, which quote a 984k context window. That is not a different model and not a downgrade. It is the thinking-mode input ceiling, 983,616 rounded, and it is the honest number to size against if you call the model with reasoning on — which is the default.
The effort remap is where an OpenAI-shaped client loses a level
Qwen4-Max exposes three reasoning levels, not five: low, high and xhigh, defaulting to high. Call it through an OpenAI-compatible client and an out-of-range effort string is folded rather than rejected — Alibaba's table maps minimal to low, medium to high, and max to xhigh. Through the Anthropic Messages format, medium becomes high and max becomes xhigh.
Both mappings are defensible and both are silent. A pipeline tuned against a five-level model that requests medium here receives high-effort reasoning: more tokens, more latency, more spend, and no error to tell you it happened. A pipeline that requests max expecting the top of a five-rung ladder gets xhigh, which happens to be the top of this one — so the output looks right and only the invoice disagrees. The fix is one line per route: pin the level you actually measured on, in the vendor's vocabulary, and assert on it in staging. Anything that passes an unmapped effort string straight through to a model with a shorter ladder will drift, and it will drift in the direction of spending more money.

What the launch numbers say, and what they do not
Alibaba's launch evaluation, vendor-reported and not independently reproduced, puts Qwen4-Max at 52.69 on its Work360 benchmark against 47.74 for Qwen3.8-Max, with average cost per task falling from $4.43 to $1.67 and output token use down 24% at xhigh effort and 35% at high. On Terminal-Bench 4.0 the vendor reports 48% and 49% fewer output tokens at the same two settings, and function-level pass rates on its secure-coding benchmark rise from 22.0% to 31.8%. Those are the vendor's case for the architecture: a 2.4-trillion-parameter sparse mixture-of-experts activating 71 billion parameters per token, plus 328 billion of N-gram embedding parameters, with three Gated DeltaNet layers per sparse-attention layer.
The one outside reading available on 17 November is Artificial Analysis, which scores Qwen4-Max at 43 on Intelligence Index v5.0.1 and ranks it 38th of 81 models on that index. It also measures $3.27 of compute per index task, held up by 127 million output tokens over the run against a median of 78.6 million, and it records no speed measurement for the model at all — no time per task, no output rate. A 43 beside a vendor claim of beating the frontier is the tension the next month of coverage will be about. It is also the only number in this article Alibaba did not produce.

If you were going to call it this week
OrcaRouter does not route Qwen4-Max, and it is worth saying that plainly rather than implying otherwise. The catalogue carries Qwen3.8-Max — the model Qwen4-Max replaces — at $2.00 in and $6.00 out per million with the vendor's list price passed through at 0% markup, alongside Claude Opus 5.5, GPT-6.1 Sol, Grok 4.5 and Tencent HY4 Preview. It does not yet carry the Qwen4 generation.
The honest arrangement for a team evaluating Qwen4-Max today is therefore a direct Alibaba key for the new model plus a routing layer for everything you would fall back to. That is less awkward than it sounds, because the thing that makes a new model expensive to adopt is not the token price — it is that trying it means a second contract, a second SDK and a second set of retry semantics sitting on a production path. Keeping the incumbent behind one endpoint with automatic failover across providers lets the Qwen4-Max experiment run beside a path that already works, and lets the comparison happen on your own traffic rather than on Alibaba's benchmark. When the model does reach the catalogue, the price you see will be the vendor's own, live the day it changes, because that is what pass-through means: the Global card at $1.413 and $4.24 would appear as $1.413 and $4.24, not at a reseller's spread.
What to watch
Alibaba's launch post closes with a line that matters more than any benchmark in it: Max is only the first look at Qwen4, and smaller models in the family are coming to open source. That is a promise about the tier below, not about this one, and it does not change the fact that Qwen4-Max itself is a model you rent from a single vendor and from platforms that rent it from the same vendor. In that market the documentation is the product differentiator, and the documentation says 983,616 in and 131,072 out — not one million of anything.
Two dates belong in a calendar. The 25 November retirement of the preview identifiers, for anyone whose config still says qwen4-max-preview. And whatever Alibaba publishes next about the thinking budget, because a 262,144-token shared pool on a model marketed as 1M-context is the single most likely number to be revised once real agent workloads run into it.
