Qwen4-Max Preview vs Claude Opus 5.5: A Fourteen-Day Model vs the One You Can Still Call in December

A​libaba opened Qwen4-Max Preview at 10:00 on 11 November 2026 and, on the same documentation page, said when it closes: 24:00 on 25 November, UTC+8. Fourteen days. A​nthropic released Claude Opus 5.5 on 22 September 2026 at $4.00 per million input tokens and $20.00 per million output, with a retirement commitment of not sooner than 22 September 2027. Those two sentences already settle most of this comparison, because one of these models is a fortnight-long experiment and the other is a line item you can put in next year's budget.

What the dates do not settle is whether the experiment is worth running. There is no benchmark for Qwen4-Max Preview — not a vendor table, not an independent run, not a single published score — so any page that ranks it against Claude Opus 5.5 on quality is guessing. This one does not. It compares the four things you can actually verify before Friday: what each model accepts as input, how much output a single response can hold, what a task costs, and how long you are allowed to keep using it.

What each one is, in the form you can call today

A​libaba spells the product Qwen4-Max-Preview in its announcement and the API string qwen4-max-preview in its documentation; there is also an alias, qwen4.0-max-preview, and A​libaba notes that both work but that billing renders under the 4.0 form. Neither name resolves to a public model card, a parameter count, or a licence, because there is nothing to download.

Claude Opus 5.5 is the opposite kind of object: a documented, priced, multi-platform endpoint whose every limit is on a page you can cite. Put the two side by side and the contrast is not really about capability at all.

• Context window — Qwen4-Max Preview 1,000,000 tokens vs Claude Opus 5.5 1,000,000 tokens. A genuine tie, and the one dimension where the preview needs no caveat.

• Maximum output per response — 393,216 tokens for the preview, a figure A​libaba states is shared by thinking and the final answer vs 128,000 tokens for Claude Opus 5.5, or 300,000 on the Batches API with a beta header.

• Input — text, images and video vs text and images. This is the only axis where one model simply cannot do the job the other can.

• Price per million tokens — not published; metered in Token Plan Credits against a subscription vs $4.00 input and $20.00 output, with cache reads at $0.20 and 5-minute cache writes at $5.00.

• Thinking default — reasoning_effort on, defaulting to xhigh, with low and medium available vs adaptive thinking always on, effort defaulting to medium on a low-to-max scale. The defaults point in opposite directions, which matters more than it sounds.

• Published benchmark — none vs Artificial Analysis Intelligence Index 59, ranked fifth of 74 models on that measure.

• How you reach it — Token Plan subscription only, on the Singapore base URL token-plan.ap-southeast-1.maas.aliyuncs.com, plus the Q​wen app, Q​wen Chat, Agent Mode, Qoder, QwenWork and QwenStudio vs the Claude API, Amazon Bedrock, G​oogle Cloud, Microsoft Foundry and Claude Platform on AWS.

• Availability window — 11 to 25 November 2026 vs active, with retirement not sooner than 22 September 2027.

The asymmetry in that list has a practical consequence that is easy to skip past. Qwen4-Max Preview has two doors — a product door with no key required, and an API door that needs a Token Plan subscription and a base URL you do not already have — while Claude Opus 5.5 is reachable from any O​penAI-shaped client you already own. If your evaluation is "does this model understand my codebase", the chat door is enough. If it is "does this survive contact with my pipeline", you are provisioning a second contract for a model that expires in two weeks.

The 393K output ceiling is where the documentation hides a trap

393,216 is the number every summary of this launch leads with, and it is a real advantage: it is just over three times Claude Opus 5.5's 128,000-token ceiling. Whole-repository rewrites, long transcript synthesis, single-response generation of an entire small application — the preview is built for outputs that do not fit in most models' response budget.

Read the footnote, though. A​libaba says the 393,216 tokens are shared by thinking and the final answer, and then adds a sentence that will break ports: starting with this model, max_tokens in O​penAI Chat Completions is equivalent to max_completion_tokens, and both cap the combined thinking-plus-answer spend. Every earlier Q​wen model kept the older behaviour, where the cap applied to the answer. So a request lifted from qwen3.8-max with max_tokens: 8192 now buys 8,192 tokens of thinking and whatever is left — which, at the default xhigh effort on a hard prompt, can be nothing. The failure mode is a truncated answer, not an error you can catch.

Fairness requires one correction to the way this is usually reported. A​nthropic's max_tokens also bounds thinking and text together within a single response, so a shared budget is not a Q​wen invention and is not a reason to prefer one vendor. What is different is the size of the budget, and the fact that Q​wen changed the meaning of a parameter its own users had already learned. That is a migration cost, and it lands on you, not on the vendor.

The second trap is the default effort. Qwen4-Max Preview reasons at xhigh unless told otherwise; Claude Opus 5.5 reasons at medium. Two unconfigured calls, compared side by side, are not measuring the same amount of work — and on the preview, a deeper default eats a larger share of the response budget before the first word of the answer exists.

Video input is the one gap with no workaround

Claude Opus 5.5 accepts text and images. Qwen4-Max Preview accepts text, images and video, and A​libaba names the use cases directly: screenshot analysis, document understanding, video analysis.

If your pipeline already samples frames and sends them as images, nothing changes — both models can read the frames, and the difference collapses into token accounting. If your pipeline wants to hand over a file and ask what happened in it, only one of these models can accept the question. Temporal reasoning over a supplied clip is not something you can approximate by describing the clip in text; you can reconstruct an approximation, and you will spend a researcher's week discovering how much of the answer lived in the motion.

That is the cleanest decision rule in this comparison, and it does not require a benchmark. A team with video in the input path has one candidate today. A team without it has no reason to accept a fourteen-day expiry as part of the deal.

The independent numbers, and the one that is not about this model

Artificial Analysis has no page for Qwen4-Max Preview. As of 11 November there is no measured speed, no cost-per-task, no index score and no provider count for it — the preview is too new and, at fourteen days, may never get an entry. So the only honest way to put the two on one chart is to use the most recent Q​wen flagship that has been measured, Qwen3.8-Max, and say plainly that it is a proxy.

• Intelligence Index — Claude Opus 5.5 59, ranked fifth of 74 vs Qwen3.8-Max 34, below the median of 40 for its price tier. Per Artificial Analysis.

• Output speed — 115 tokens per second vs 53. Per Artificial Analysis, same harness.

• Time to first answer token — 353.75 seconds vs 122.57 seconds. Opus 5.5 is the slower model to start and the faster one to finish.

• Time per task — 25.0 minutes vs 54.1 minutes, against a median of 12.6.

• Cost per task — $10.04 vs $10.02. Read that twice; it is the most interesting number on this page and the next section is about why.

• Verbosity — 218 million output tokens on the index vs 150 million, against a median of 72.8 million.

There is one measurement that does touch a Q​wen 4 checkpoint against Claude Opus 5.5 directly, and it deserves care rather than a headline. The GameFeel Benchmark — a creative and game-development eval that scores build, bug-fix, game feel, ideas, visuals and audio — carries a row for Q​wen-4-Flash-Early averaging 49.9 against Claude Opus 5.5 at XHigh effort averaging 47.2. Four things must be said about that before anyone quotes it. It is a Flash-tier checkpoint, not Max. The -Early label is the benchmark author's guess at which unreleased tier an unlabelled endpoint belongs to, not an A​libaba confirmation. The two rows were not billed the same way, so the $0.40 and $2.60 per-task figures are not like-for-like. And the average hides the split: Opus wins on bug-fixing, 65 to 63, and on reverse-evaluation, 21.67 to 19.95, while the Q​wen checkpoint wins on ideas 54 to 41, visuals 65 to 63, audio 30 to 25 and game feel 52 to 50. Bug-fixing and reverse-evaluation are the two subtasks closest to production work; ideas and audio are the two furthest from it.

The same bench also shows what effort does to Opus 5.5: at Medium it scores 42.3 for $0.88 a task, against 47.2 for $2.60 at XHigh. Nearly three times the spend for five points. If you are running Opus at maximum effort by default, that is the line to look at.

What a task costs when you cannot price one side

Start with the half that has a rate card. A job that sends 200,000 input tokens and returns 20,000 output tokens costs $0.80 in input and $0.40 in output on Claude Opus 5.5 — $1.20. Reuse that input from cache and the input side falls to $0.04, which makes the same job $0.44. Batch it and the whole thing halves again. Those are arithmetic, not estimates, and you can put them in a forecast.

Now the other half. A​libaba publishes no credits-per-token rate for Qwen4-Max Preview. Token Plan Personal Edition is a Credits subscription — A​libaba's four personal tiers run from $6 a month for 20,000 Credits to $43 for 200,000 — and the vendor says the rate at which a call consumes Credits is set dynamically by model type, token usage, thinking mode and tool calls. The only place to observe the real rate is the usage detail in the console after you have spent some. During the preview there is a promotional discount, which makes the observed rate a poor guide to the post-preview one.

So the pricing comparison is not "which is cheaper". It is "one of these has a unit price and the other has a subscription you cannot yet convert into one." And the numbers above suggest the rate card matters less than it looks: Qwen3.8-Max lists at $2.00 and $6.00 per million tokens — a third of Opus 5.5's input price and less than a third of its output price — and still lands at $10.02 per Intelligence Index task against Opus's $10.04. The cheap model generates so many more tokens per task that the rate card is almost entirely cancelled out. A cheaper list price buys you a cheaper invoice only if the model is also concise, and on every measurement available, this family is not.

Running them side by side without pretending it is one contract

Claude Opus 5.5 is on our catalogue as anthropic/claude-opus-5.5, at Anthropic's published $4.00 and $20.00 per million tokens with 0% markup, so it is one key and one bill alongside more than 200 other models, with a fallback chain and the routing DSL available if you want to send the easy half of a workload somewhere cheaper than Opus without changing your code.

Qwen4-Max Preview is not on our catalogue, and it is worth being blunt about what that means: we do not serve it, and you cannot put it in a failover chain here. Reaching it means a Token Plan API key, the Singapore base URL, and a second contract — which is exactly the sort of thing a two-week model rarely justifies building. If you want to evaluate it, run it as an offline bake-off on your own key against prompts you already know the answer to, and keep production pointed at something that will still answer requests on 26 November.

What the preview is good for is the question it forces. Two weeks is long enough to find out whether video-native input and a 393K response budget change what your product can do — and short enough that you will not accidentally end up depending on the answer.

Who should do which, this week

If you have video in the input path and no production dependency on the result, open Qwen4-Max Preview through Q​wen Chat today, spend an afternoon on the clips you actually care about, and treat whatever you learn as research rather than architecture. Set reasoning_effort to medium for the first pass so the answer budget is not consumed before the answer starts, and set max_tokens high enough that thinking and output both fit.

If you are shipping anything that has to work in December, Claude Opus 5.5 is the only one of the two with a price you can forecast, a benchmark you can cite to a client, and a date on which it will still exist. Its weaknesses are documented rather than mysterious — it is slow to first token, verbose, and expensive per task at high effort — and all three are tunable through the effort parameter.

And if you want both, in that order, the sequencing is the point: prove the capability on the free door, build on the model that has a rate card, and revisit the question in December when the preview's successor arrives with, one assumes, a price and a benchmark attached.