A generated hero title card reading 'Qwen4-Max vs Qwen3.8-Max-Prime' with the subtitle 'The fast lane costs double and finishes last.' and two price-and-throughput panels: Qwen4-Max at '$1.50 / $4.50 per M' and 104,940 tokens per task in 32.1 minutes, against Qwen3.8-Max-Prime at '$3.301/$9.902 per M' and 170,794 tokens per task in 47.5 minutes. Flat vector editorial styling on a white background with a blue-to-cyan gradient wash, and the OrcaRouter logo composited bottom right.
Guides & Insights

Qwen4-Max vs Qwen3.8-Max-Prime: the fast lane costs double and finishes last

Author

jinhao.song

Date Published

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

Qwen3.8-Max-Prime exists for exactly one reason: to charge roughly double so the tokens arrive sooner. Qwen4-Max, generally available since 16 November 2026, undercuts the model Prime was built on — $1.50 in and $4.50 out per million tokens in Singapore against $2.00 and $6.00 — and it finishes a task in about two-thirds the wall-clock time. So the comparison a reader actually needs is not "new flagship versus old flagship". It is whether the premium speed lane Alibaba sells on top of Qwen3.8-Max still buys anything you cannot get by simply moving to Qwen4-Max at the standard rate, and the honest answer is narrower than either vendor page suggests: Prime still streams faster per token, and it is no longer clearly faster per job.

There is a second, quieter complication. As of 18 November the Prime model ID does not appear on Alibaba's Fast mode documentation page or on its model rate card in either the international or the China edition. That page was last updated on 11 October 2026, and it now lists three Prime IDs, all of them GLM or Wan — no Qwen variant at all. Absence is not the same as withdrawal, and Alibaba has not published a retirement notice we can find, so treat this as "no longer listed" rather than "discontinued". But if you are budgeting a switch this week, it changes the question from which of the two you should buy to whether the older one is still on the shelf.

One of these is a model, the other is a serving lane

The distinction decides the whole comparison, and the marketing does not draw it.

• Qwen4-Max — a new flagship from Alibaba's Qwen team, GA 16 November 2026, metered per token in six regions, 1,000,000-token context, text, image and video input with text output, thinking on by default at reasoning_effort=high.

• Qwen3.8-Max-Prime — not a model. It is Qwen3.8-Max, the 2.4-trillion-parameter 3.8 flagship that reached GA on 3 August 2026, served through what Alibaba calls Fast mode (Prime mode, 优速模式). Same weights, same context, same tool surface, same usage limits. Alibaba's own documentation says the capabilities and restrictions are "the same as the original model" and the only thing you change is the model parameter.

That matters because a serving lane ages differently from a model. Prime inherits every capability ceiling of the 3.8 generation and every one of its token-efficiency characteristics; what it adds is throughput headroom on Alibaba's own infrastructure, priced at a multiplier. When the generation underneath it is replaced by something cheaper and more token-efficient, the lane keeps its speed and loses its reason.

The price sheet, and the inversion sitting in it

These are list prices from Alibaba's rate card, not measured bills. Per million tokens, input and output:

• Qwen4-Max, Singapore — $1.50 / $4.50, flat across the full 1,000,000-token window.

• Qwen4-Max, everywhere else — $1.413 / $4.24 in Beijing, Hong Kong, Frankfurt, Virginia and Tokyo.

• Qwen3.8-Max, Singapore — $2.00 / $6.00; $1.650 / $4.951 in the Global regions.

• Qwen3.8-Max-Prime — $3.301 / $9.902 on the international card, with cache hits at $0.413. That is exactly 2.0× the standard ID's Global rate, not a rounded approximation. We read those figures off the rate card in late September; they are the last published Prime numbers we have, and the row is no longer on the page today.

A generated two-column scoreboard titled 'Qwen4-Max vs Qwen3.8-Max-Prime — the scoreboard'. Left column 'Qwen4-Max' reads Intelligence Index 42.9, HLE Diamond 44.9%, APEX Agents 59.1%, Terminal-Bench 4.0 48.5%, 104,940 tokens per task, $3.27 per task. Right column 'Qwen3.8-Max-Prime' reads Intelligence Index 33.7, HLE Diamond 24.0%, APEX Agents 42.4%, Terminal-Bench 4.0 38.9%, 170,794 tokens per task, $10.02 per task. Footer reads 'Scores: Artificial Analysis Intelligence Index v5.0.1. Prime mode runs Qwen3.8-Max weights, so its scores are Qwen3.8-Max's.'

• Cache-read rate — $0.075 per million in Singapore and $0.071 elsewhere, against Prime's $0.413 cache hit. Under a fifth of the price, on the line a long agent loop pays most often.

Line the first and fourth up and the inversion is stark: the two-day-old flagship costs about 43% of what the fast lane of the previous generation costs on input, and 43% on output. Even against the plain Qwen3.8-Max it replaces, Qwen4-Max is 25% cheaper in Singapore — and the vendor is positioning that as a price cut, which it is, but only if the new model does not spend the savings on more tokens. More on that below, because it partly does and partly does not.

The cache line is the one that compounds. A long agent loop with a fixed system prompt and tool schema re-reads its prefix on every turn, so the cache-read rate, not the headline input rate, sets the marginal cost of turn fifty. At $0.075 against $0.413, Qwen4-Max re-reads a cached prefix for about a fifth of what Prime charges — a gap 2.4 times wider than the 2.3× on fresh input, and the one that decides whether a 400,000-token document sweep is affordable at all.

What Prime actually buys: tokens per second, not tasks per hour

Alibaba's Fast mode page states the benefit plainly: TPS is raised to 1.5 to 2 times that of the standard API, aimed at AI coding assistants, multi-step agent reasoning and real-time conversation. It also describes an unusual throttling rule — when you hit the rate limit, if the platform still has spare capacity you are not throttled, so the throughput actually available will not fall below the stated limit. That is a floor, not a cap, and for an interactive coding assistant the tail matters more than the mean.

The number has never been independently measured. What has been measured is the standard Qwen3.8-Max build: Artificial Analysis records 60.0 output tokens per second on Alibaba's own API, with 3.6 seconds to first token. Qwen4-Max, on the same board, records 54.5 tokens per second and 6.9 seconds to first token. So if Prime really delivers the top of its stated range, it streams at roughly 120 tokens per second — more than twice Qwen4-Max — and its first token arrives sooner.

Then read the other half of the same table, and the advantage inverts:

• Output tokens per index task — Qwen4-Max 104,940 vs Qwen3.8-Max 170,794. The new model spends 38.6% fewer tokens getting through the same evaluation suite, at the same xhigh effort setting on both sides.

• Wall-clock per task — 32.1 minutes vs 47.5 minutes. Qwen4-Max finishes a third sooner even though it streams slower.

• Cost per task — $3.27 vs $10.02, on published list prices. Multiply that $10.02 by Prime's flat 2.0× premium and the same task lands near $20.

That last figure is arithmetic on the rate card, not a measurement — Artificial Analysis timed the standard ID, not the Prime lane, and nobody has published a Prime-mode benchmark we can find. State it as an estimate. But the direction is not in doubt: if you pay 2.0× per token for a lane that emits the same tokens, you pay 2.0× for the task. Prime makes each token faster and each job more expensive; Qwen4-Max makes each token slower and each job cheaper and shorter.

For a chat interface where a human is watching characters appear, that difference is real and Prime still wins it. For an agent that runs twelve sequential calls and bills you at the end, the token count is the wall clock.

The capability gap, measured rather than claimed

Alibaba's launch material for Qwen4-Max quotes token-efficiency improvements of roughly 24% at xhigh effort and 35% at high on its internal Work360 benchmark, plus a 62% drop in average cost per task. Those are vendor-reported and unreproduced. The independent board tells a consistent but less dramatic story, and it is the one to plan against.

On Artificial Analysis Intelligence Index v5.0.1, which both models are scored under at the same effort setting:

• Index — Qwen4-Max 42.9 (the site rounds it to 43) vs Qwen3.8-Max at 33.7. A nine-point generation gap, and the same board records the 3.8 row as deprecated, superseded by Qwen4-Max.

• Humanity's Last Exam, Diamond — 44.9% vs 24.0%. This is the widest single divergence in the row and the one that most changes what you can delegate.

• APEX Agents — 59.1% vs 42.4%. Agentic task completion moves almost seventeen points, which is the difference between a workflow you supervise and one you check afterwards.

• Terminal-Bench 4.0 — 48.5% vs 38.9%. A real gain, and worth reading next to the 88.8% the 3.8 build records on the older Terminal-Bench 2.1, which Qwen4-Max has not been scored on. Do not compare across those two versions.

• FrontierCode 2.0 — 26.5% vs 23.4%. The smallest gap on the board. If your workload is code generation against a repository, the generational jump is three points, not nine, and Prime's throughput may still matter more to you than the index.

• Omniscience — 27.7 vs 10.5, with hallucination rates of 28.4% and 27.1%. Qwen4-Max knows substantially more and confabulates at about the same rate. Those two facts together are why it is safer on open-ended research and no safer on anything you cannot check.

Two gaps cut the other way, and both are about what you can pin. Qwen3.8-Max has a dated snapshot, qwen3.8-max-0902, still on the rate card at the same price, so you can freeze behaviour across a release. Qwen4-Max has no snapshot ID, no fine-tuning and no dedicated deployment — Alibaba's page marks both as unsupported. If your compliance story depends on a model that cannot move underneath you, the older generation is the only one of the two that offers it, and Prime inherits that same dated build.

The availability question nobody has priced in

Here is the part that makes this comparison urgent rather than academic. Check the Fast mode page today and you will find three supported IDs — glm-5.3-prime, glm-5.2-fast-preview and wan3.0-video-prime — in Beijing and Singapore. No Qwen model is listed, in either region. Check the model rate card in the international or the China edition and there is no qwen3.8-max-prime row; the Qwen entries run qwen4-max, qwen3.8-max, qwen3.8-max-0902 and down the 3.7 line. And there is no qwen4-max-prime anywhere — the new flagship has no fast-lane variant at all.

We cannot tell from the outside whether Alibaba retired the Qwen Prime ID, moved it into a regional product page we have not found, or simply dropped it from a documentation table during a restructure. What we can say is narrower and still useful: a reader planning a latency-sensitive build on Qwen3.8-Max-Prime this week should confirm the ID still resolves on their own workspace before designing around it, because the public documentation no longer promises it. That is a different risk profile from September, when the ID sat on the rate card next to the standard one.

It also reframes the decision. If Prime is still available on your account, you are choosing between a fast lane on an older, weaker, more expensive model and a slower lane on a newer, cheaper, stronger one. If it is not, the choice has already been made for you, and the only remaining question is how long you keep paying $2.00/$6.00 for the standard 3.8 ID instead of $1.50/$4.50.

A screenshot of Alibaba Cloud's Fast mode documentation page, captured 18 November 2026 in English, showing the supported model table with glm-5.2-fast-preview and the other GLM and Wan Prime IDs listed for Beijing and Singapore, and no Qwen row anywhere on the page.

Who should pick which

• Interactive coding assistant, human watching the stream — if Prime still resolves on your workspace and the UX is measured in characters per second, the 2× throughput is the product. Qwen4-Max at 54.5 tokens per second is not slow, but it is not a fast lane, and its 6.9-second first token is nearly double the standard 3.8 build's.

• Agentic pipelines, batch evaluation, document sweeps — Qwen4-Max, without much argument. Fewer tokens per task, fewer minutes per task, a third of the cost per task, and a cache-read rate more than five times cheaper than Prime's on long prefixes.

• Workloads that must be pinned — qwen3.8-max-0902, the dated 3.8 snapshot, at standard rates. Neither Prime nor Qwen4-Max gives you a frozen checkpoint.

• Code generation against a large repository — the closest call on the board. Three index points of FrontierCode 2.0 improvement may not be worth losing a throughput lane you already have instrumented, particularly if your agent is the bottleneck rather than the model.

• Anyone who has not migrated yet — run the index-suite arithmetic on your own traces before you commit. The 38.6% token reduction Artificial Analysis measured is an average over a mixed evaluation; your workload may compress better or not at all.

Wiring the choice into one endpoint

Neither Qwen4-Max nor Qwen3.8-Max-Prime is in OrcaRouter's catalogue today — we checked both model pages and both return nothing, and this piece is not going to imply otherwise. What is there is the model this whole comparison is built on: qwen/qwen3.8-max and its dated snapshot qwen/qwen3.8-max-0902, served at the provider's list price with 0% markup, which today means $2.00 in and $6.00 out per million tokens. OrcaRouter's own page for that ID reports a p50 time-to-first-token of 4.34 seconds and a p95 of 10.00 seconds over the trailing week — our measurement, on our traffic, not the vendor's.

That matters for the decision in a specific way. Alibaba revised Qwen4-Max's price twice between the preview on 11 November and GA on the 16th, and the Global card is already 5.8% below the Singapore one. Whatever you build, the rate will move under you. A pass-through gateway means a vendor cut lands on your bill the same day rather than at the next contract renewal, and one key covers both generations, so the migration from qwen3.8-max to qwen4-max is a model string rather than a second integration. Until Qwen4-Max is routable here, the useful pattern is the boring one: keep the standard 3.8 ID on the key you already have, evaluate Qwen4-Max directly against Alibaba while its commercial terms settle, and let automatic failover cover the gap in either direction.

What would change this verdict

Three things, and each is checkable in the next few weeks. A Qwen4-Max Prime variant would restore the fast lane on the current generation and make this a throughput-versus-throughput comparison again, at 2× the $1.50/$4.50 rate. An independent Prime-mode measurement would replace the vendor's 1.5–2× claim with something you can put in a capacity plan — nobody has published one, and until somebody does, the multiplier is a promise. And a snapshot ID for Qwen4-Max would remove the only real argument left for staying on the 3.8 line, which right now is not price, not capability, and not speed on a full task, but the ability to freeze a checkpoint.

For everyone else the answer is already settled by the arithmetic: the fast lane costs more per token, more per task and more per cached re-read than the model that replaced it, and it does less. That is a strange thing for a premium tier to be, and it is what happens when a serving lane outlives the generation it was priced against.

A screenshot of the Artificial Analysis page for Qwen4 Max, captured 18 November 2026, showing an Intelligence Index of 43 and a Cost per Intelligence Index task of $3.27 alongside the surrounding comparison charts.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily