Qwen4-Max vs Gemini 4 Argon: One Column Is Empty, and the Other Costs $9.95 a Task

A matchup between Qwen4-Max and Gemini 4 Argon looks, on a rate card, like a coin flip: neither vendor will sell you tokens for either model at a published per-token rate today, and both arrive with a large context window and a preview-shaped caveat. It stops looking like a coin flip the moment you try to run them. Gemini 4 Argon has an identifier, an endpoint and a metered price you can be billed against. Qwen4-Max has a subscription, a Singapore-only base URL and a retirement date fifteen days after it launched. One of these is a model you can put in a request this afternoon; the other is a model you can put in a request this afternoon and not next month. That asymmetry — not a benchmark delta — is the decision this page is about, because a benchmark delta does not exist yet.

What each one actually is, dated

Gemini 4 Argon is G​oogle's frontier model, announced on 30 September 2026 in a DeepMind post credited to Koray Kavukcuoglu, and rolled out in three stages that began with a named cohort of cyber defenders in the Fairwind Program. At announcement it had no model ID in the G​emini API and no endpoint; what it had was a nineteen-row vendor benchmark table and a five-page eval methodology. Since then it has acquired the thing that matters operationally: it is in OrcaRouter's catalogue as google/gemini-4-argon, at G​oogle's list price of $2.00 per million input tokens and $10.00 per million output, and it has been moving real traffic there — 41.8 million tokens over the seven days to 11 November 2026.

Qwen4-Max is the flagship tier of the Q​wen 4 family A​libaba named from the stage at Apsara in Hangzhou on 22 September 2026. Its preview opened at 10:00 UTC+8 on 11 November 2026, model ID qwen4-max-preview, alias qwen4.0-max-preview, and it will be retired after 24:00 on 25 November 2026. There is no pay-as-you-go access: it is served through A​libaba's Token Plan subscription only, in the Singapore region, on a dedicated base URL. No per-token price, no parameter count, no benchmark table, no weights.

So the comparison has a shape worth naming precisely. Both are previews in the ordinary sense — neither is a settled, generally available product. But Gemini 4 Argon's preview is a rollout stage that is widening, and Qwen4-Max's preview is a fifteen-day window with a stated end. Preview as a phase versus preview as an event.

The empty column, and how empty it is

Every comparison of these two eventually needs the row that says who is smarter, and here is that row:

• Vendor benchmarks — Gemini 4 Argon: a nineteen-row table, G​oogle's own harness, competitor columns taken from competitors' self-reported figures. Qwen4-Max: none.

• Independent score — Gemini 4 Argon: 50 on the Artificial Analysis Intelligence Index under its current v5.0.1 revision, measured at high effort. Qwen4-Max: not measured; the leaderboard carries no Qwen4 entry.

• Context window — Gemini 4 Argon: 1,000,000 tokens. Qwen4-Max: 1,000,000 tokens. The only row where they are level, and it is level because both published a round number.

• Output ceiling — Gemini 4 Argon: 65,536 tokens in the standard path, with long-decode generation extending up to 1M. Qwen4-Max: 393,216 tokens shared between thinking and the answer.

• Input modality — Gemini 4 Argon: text, image, audio, video and PDF. Qwen4-Max: text, image and video. No audio, no PDF path documented.

• Price per million tokens — Gemini 4 Argon: $2.00 in, $10.00 out, $0.10 cached input. Qwen4-Max: not published; metered in Credits against a subscription.

• Access — Gemini 4 Argon: metered API, callable now. Qwen4-Max: Token Plan subscription, Singapore only, interactive use only, ends 25 November.

• Independent cost per task — Gemini 4 Argon: $9.95. Qwen4-Max: not measured.

Seven of eight rows have a value on the right. The row that decides "which is better" has nothing in it, and no amount of reading will fill it. That is the honest state of this matchup on 11 November 2026, and it is worth more to state plainly than to paper over with a projection.

Argon's rate card and Argon's bill are two different numbers

The most useful number in the Argon column is not its index score. It is $9.95 — the cost to complete one Intelligence Index task — sitting next to a rate card of $2.00 and $10.00. On the index suite, the average task took Argon 29.0 minutes and produced enough tokens that, at a $10 output rate, the average task costs roughly five times what the rate card suggests a task should cost. That is not a pricing error. It is what a high-effort reasoning model does: it writes a great deal before it stops.

The same index run puts Argon's median time to first answer token at 40.12 seconds. Read against the seven-day figure on our own model page — a p50 time to first token of 10.00 seconds on real OrcaRouter traffic — the two numbers are measuring different things, and the gap between them is a fair illustration of how much of a benchmark's latency is the harness rather than the model. Neither number is wrong; quoting one as though it were the other is.

For anyone budgeting a migration, the practical reading is this: a $10 output rate on a model that reasons at high effort is a rate card, not a forecast. If your workload resembles the index suite — long agentic tasks where the model writes a lot before answering — the per-task cost will run well above what a token estimate says. If it resembles a chat product, it will not. The only way to know which you have is to measure your own traffic, which is a strange thing to have to say in a comparison and is nonetheless true of every reasoning model released this year.

Why the Q​wen side has a countdown instead of a price

A​libaba's choice to ship Qwen4-Max through a subscription rather than a metered endpoint tells you something about what the preview is for. Token Plan Personal Edition is explicitly scoped to interactive use in coding and agent tools and explicitly not for automated batch scripts, and it is confined to one region. A company that intended to compete for production traffic on a flagship would not put it behind a $6-a-month consumer subscription with four tiers and a 20,000-credit floor. A company demonstrating a capability, gathering feedback under a freeze, and keeping its metering options open would.

The freeze is the part teams underestimate. A​libaba's page states that no changes will be made during the preview window. Whatever the model does wrong on day one, it will still do on day fourteen. That is a defensible engineering decision and a poor foundation for a dependency, because it means the only feedback loop available to you is a written report to a vendor who has already said it will not act on it until the window closes.

There is also a token-accounting change that lands with this model and only this model: max_tokens now caps thinking plus answer together, not the answer alone. Every earlier Q​wen model keeps the old behaviour. A request lifted from Qwen3.8-Max with a 4,096-token cap will now spend part of that budget reasoning at xhigh effort before it writes anything, and the symptom is a truncated response rather than an error.

Calling them, honestly

Gemini 4 Argon is one line from an O​penAI-shaped client on our key, at G​oogle's rate with nothing added, with an A​nthropic-format endpoint as well and zero-data-retention available for requests that need it. Qwen4-Max is not on our catalogue — we do not serve it, and we will not pretend otherwise; the catalogue returns 217 model identifiers today and not one of them is a Q​wen 4 tier.

That gap is exactly where routing earns its place rather than being a slogan. If you are evaluating the Q​wen preview, put it behind something you already trust in a fallback chain, so a timeout or a truncated answer degrades to a working model instead of failing the request. If you are running Argon on a production path, the same mechanism covers the other risk: a 10-second p50 time to first token is a real latency floor, and a chain that fails over when it is exceeded costs you a config line rather than an incident.

The cost argument for having both behind one key is mundane and worth stating: two vendors, two billing systems and two SDK integrations become one key and one client. The preview's Credits subscription does not become metered by being behind our gateway — we cannot resell what A​libaba will not sell — so the honest version of the claim is that Argon is one key away and Qwen4-Max is one subscription away, and those are different amounts of work.

The decision, given that one column is empty

If you need a frontier model with a million-token window and video input this month, the choice is not between these two. It is Gemini 4 Argon, on the evidence that it is callable, metered, measured at 50 on the current index revision, and shipping traffic — with the caveats that its $2/$10 card understates a high-effort bill and its vendor benchmark table compares G​oogle's harness against competitors' press releases.

If you are deciding whether to spend an afternoon on the Q​wen preview, the answer depends on one question: does your workload involve very long outputs from very long inputs? A 393,216-token response budget on a 1M context is the largest published output ceiling of any model in this comparison, and if that is the constraint you keep hitting, fifteen days is enough time to find out whether Qwen4-Max handles it. If your workload is ordinary chat or code completion, the preview teaches you little that Qwen3.8-Max cannot, and it teaches it in a currency you cannot convert to dollars.

What would change this page is a Q​wen 4 model card with a score on it. When that lands, the empty row fills and the comparison becomes the one readers expect. Until then the useful comparison is a callable model against a fifteen-day experiment, and the only person who can say which one you need is the one who knows whether the answer has to still be there in December.