A generated 16:9 title card on a dark navy background reading 'Qwen4-Max' with the subtitle 'Two dates, one scoreboard', and three rounded cards: MODEL DATE 16 Nov 2026, SECOND SELLER 18 Nov 2026, INDEX SCORE 43. A footer line reads 'Artificial Analysis Intelligence Index v5.0.1'. OrcaRouter logo bottom-right.
Guides & Insights

Qwen4-Max Is on Sale in Two Places Now. The Independent Scoreboard Is the Part Worth Reading

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Qwen4-Max went on general sale on 16 November 2026, and by the morning of 18 November a third-party inference platform was advertising it too, with the same 1M-token context and the same text, image and video inputs. That is the news, and it is a small piece of news: a second seller is a distribution event, not a model event. The thing that arrived in the same window and has had almost none of the attention is a measurement. Artificial Analysis now carries a Qwen4-Max entry, and on its Intelligence Index it scores the model at 43, against 34 for Qwen3.8-Max, the predecessor it replaces.

Both of those numbers come with context that decides whether they are good news, and the vendor's own blog does not supply it. So this piece does three things in order: what actually shipped on 16 November, what the independent row says that the launch post does not, and what changes — and what quietly breaks — when the same model is served by somebody other than A​libaba.

Two dates, and only one of them is a release

The model's date is 16 November 2026. A​libaba's Q​wen team dated its launch post 16 November, and the Model Studio documentation page for the model carries an update timestamp of 2026-11-16 09:46:08 with the heading "qwen4-max (Latest Model)" and a plain model identifier of qwen4-max. That page is the one that matters commercially: it replaced a preview identifier, qwen4-max-preview, that had opened on 11 November on a credits plan with a published retirement date of 25 November. General availability meant a metered rate card, six regions and no expiry.

The 18 November event is different in kind. A platform that is not A​libaba began listing the model for its own customers, two days after the price card appeared. Nothing about the model changed on 18 November: no new weights, no new context figure, no new price from the lab. If you are reading coverage that frames this week as a launch, the framing is one step behind — Qwen4-Max launched on the 16th, and what happened on the 18th is that it became resellable at scale. That distinction is not pedantry. It tells you which questions to ask: not "what can it do", which was answered on Monday, but "what survives the trip through a reseller", which nobody has answered.

A generated two-column scoreboard titled 'Qwen4-Max vs Qwen3.8-Max' with five rows: index score 43 against 34; output tokens per run 127M against 150M; wall-clock per task 32.1 min against 47.5 min; cost per task $3.27 against $10.02; throughput 55 tok/s against 'not listed'. Footer reads 'Artificial Analysis Intelligence Index v5.0.1, measured November 2026'. OrcaRouter logo bottom-right.

What shipped on 16 November

The specification below is A​libaba's, from the launch post and the model documentation, and none of it has been independently reproduced — the architecture in particular is a vendor description of a training run no outsider has inspected.

• Architecture — a new Qwen4 design with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, and a further 328 billion N-gram embedding parameters

• Attention — three Gated DeltaNet layers feeding every Q​wen Sparse Attention layer, which compresses history into fixed-size states and indexes context at micro-block granularity

• Stability work — four parallel residual branches with dynamic read and write gates, described by the lab as the reason the 2.4T run had no loss spikes

• Context — 1,000,000 tokens in the vendor's documentation; Artificial Analysis measures the usable window at 984,000

• Output ceiling — 131,072 tokens with thinking off, or 262,144 shared between thinking and answer

• Modality — text, image and video in; text out. Native vision, not a bolt-on

• Reasoning effort — selectable levels, which the launch post refers to as High and Xhigh when it quotes its own efficiency figures

• Tooling — function calling, parallel tool calls, web search, structured output against JSON Schema, prefix completion, explicit and implicit caching

• Interfaces — O​penAI Chat Completions and Responses, A​nthropic Messages, and DashScope, which is the widest protocol surface A​libaba has shipped on a flagship

• What is missing — fine-tuning and model deployment are both marked unsupported. There is no checkpoint, and the launch post says only that smaller Qwen4 models are "coming to open source"

The efficiency claims are the ones the launch leans on hardest, and they are worth stating precisely because they are conditional. Against Qwen3.8-Max, the vendor reports roughly 24% fewer tokens on its Work360 benchmark at Xhigh effort and 35% fewer at High, and 48% and 49% fewer on Terminal-Bench 4.0 at the same two settings. Average cost per task on Work360 drops from $4.43 to $1.67, a 62% reduction. Every one of those numbers is vendor-reported, and every one of them is tied to an effort level — which becomes the important detail later, when a reseller decides how to expose that dial.

The independent row nobody is quoting

Artificial Analysis published Qwen4-Max on Intelligence Index v5.0.1, and the entry is unusually blunt. The score is 43, which the site describes as below average among comparable models — the median is also 43, so the model sits exactly on the line rather than above it. That is a less flattering sentence than "up nine points on its predecessor", and both are true. Qwen3.8-Max, measured as the 0902 snapshot, scores 34 on the same index.

The rest of the row is where the argument gets interesting, because it separates capability from the cost of extracting it.

• Index score — Qwen4-Max 43 versus Qwen3.8-Max (0902) 34 on v5.0.1

• Output tokens per index run — 127 million versus 150 million, against a field median of 65.1 million. Both models are extremely verbose; the new one is less so

• Wall-clock per task — 32.1 minutes versus 47.5 minutes, against a field median of 11.5. Faster than its predecessor, still among the slowest models measured

• Throughput — 55 output tokens per second, measured on A​libaba's own API

• Cost per task — $3.27 versus $10.02, which is the single most useful number in the row

Read those together and the upgrade is real but narrower than the marketing suggests. Qwen4-Max buys nine index points, cuts wall-clock by a third, and cuts the cost of finishing a task by roughly two-thirds relative to the model it replaces. It does not buy speed in absolute terms: at 32.1 minutes per task it is nearly three times slower than the median model on the same board, and it burns close to twice the median token volume to get there. If your workload is latency-sensitive, the index number is not the one that will decide your bill.

A screenshot of the Artificial Analysis page for Qwen4 Max, headed 'Qwen4 Max: Intelligence, Performance & Price Analysis' and dated November 2026, showing an Intelligence Index score of 43 against a median of 43, 55 output tokens per second, a 984k context window, $1.50 input and $4.50 output per million tokens, $3.27 per task and 127M tokens generated on the index run.

What survives the relay to a reseller

This is the part that the 18 November listing makes urgent rather than theoretical. When a model is served by a platform that is not its author, four things travel reliably and one thing does not.

The four that travel: the price card, because it is copied from the vendor's page; the context figure, because it is a headline; the modality list, because it is in the marketing copy; and the protocol shape, since most platforms expose an OpenAI-compatible endpoint regardless of what sits behind it. Those are the things every listing gets right, and they are the things every comparison article quotes.

The thing that does not travel is the effort mapping. Qwen4-Max's economics are effort-dependent by design: the vendor's own token savings swing from 24% to 35% on Work360 depending on whether you run Xhigh or High, and the documentation exposes the choice as a request parameter. A reseller that pins effort to a single default, or maps an OpenAI-style reasoning parameter onto only two of three levels, hands you a model whose cost per task can differ from the vendor's published figure by a third — with an identical price per token on the invoice. If you are benchmarking a listing against A​libaba's own numbers, pin the effort level explicitly and check what the platform does with it before you conclude anything about the price.

Two smaller things to verify: whether the 262,144-token shared thinking-and-answer budget is enforced or truncated silently, and whether prompt caching is passed through at the vendor's cache rates rather than at the full input rate. Both change real money on long agent runs, and neither is visible in a headline.

What did not change, and will not

The most consequential fact about Qwen4-Max is not on either the price card or the leaderboard: you cannot own it. There is no checkpoint for this model, and the documentation marks fine-tuning and self-deployment as unsupported. The launch post's closing line promises that smaller Qwen4 tiers will come to open source, which is a statement about the family's future, not about this model. Every route to Qwen4-Max runs through somebody's API, whether that is A​libaba's six regions or a reseller's endpoint.

That is a normal arrangement for a 2026 frontier flagship, and it has a normal consequence: your integration is a rental, and the number of landlords just went from one to more than one. Which is the only genuinely good argument for the 18 November listing — not that it is cheaper, but that it is another supply path for a model you cannot host yourself.

If you are wiring this up this week

Qwen4-Max is not in OrcaRouter's catalogue today, and this piece is not going to pretend otherwise. What is there is the model it replaces: qwen/qwen3.8-max, the 2.4-trillion-parameter flagship that scored 34 on the same index and costs $2.00 in and $6.00 out per million tokens on A​libaba's list. If you are building against the Q​wen line and want to try Qwen4-Max without betting a production path on a two-day-old commercial arrangement, the sane pattern is to keep the predecessor on the same key and route between them — OrcaRouter carries more than 200 model identifiers across providers behind one OpenAI-compatible endpoint at 0% markup, with automatic failover, so a provider-side outage or a rate limit does not take your application down with it. The routing DSL does the same thing one level down: you can compose a cheap model for the bulk of a workload and escalate to the expensive one only where the task actually needs it, rather than paying flagship rates for every call.

A screenshot of the OrcaRouter models catalogue showing the filter rail and the header '226 models, 17 providers, one API key, one bill', with a card for Google Gemini 4 Carbon and its per-million input and output rates visible.

That is not the same as having Qwen4-Max on the key, and it should not be sold as such. It is the difference between waiting for a model to reach your stack and having a stack that can absorb it when it does.

What would change this piece

Three things, in descending order of likelihood. A second independent measurement of Qwen4-Max would settle whether 43 is a floor or a ceiling — one board, one harness, one effort setting is a thin basis for a verdict, and this blog has watched index revisions move scores by double digits before. A published effort-level breakdown from Artificial Analysis, matching what it already does for GLM-5.3, would make the cost-per-task figure actually actionable rather than an average over an unspecified setting. And a Qwen4 checkpoint — any Qwen4 checkpoint — would end the rental argument entirely.

Until then the correct reading of this week is the unglamorous one. A flagship shipped on Monday with a real price card and a competitive-but-not-leading independent score; a reseller picked it up on Wednesday, which is distribution doing what distribution does; and the number that should decide your integration is not $1.50 or 43, but 32.1 minutes and a token bill that runs at roughly twice the field median. Those two are what you will actually pay.