A generated title card headed 'Endorsed for agentic coding, measured at 74th of 79', with chips reading 54 tokens per second, 32.2 minutes per index task, 179.79 seconds to first answer token and 43 on the Intelligence Index, and a footer reading 'Speed figures Artificial Analysis, read 17 November 2026'.
Guides & Insights

Qwen4-Max Is Now Being Pitched for Agentic Coding. The Independent Speed Row Says 74th of 79

Author

jinhao.song

Date Published

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

At 09:31 UTC on 17 November 2026, the official Qwen4-Max account on X posted five words and an invitation: the model is now live on a third-party platform, come try agentic coding. It was a quote-tweet of that platform's own announcement, one day after Alibaba moved the model from preview to metered general availability. The post is small and it is easy to scroll past. It is also the first time the vendor has pointed its own followers at a specific outside route for its flagship, and it picked the workload — long-horizon agentic coding — on which Qwen4-Max currently has the weakest independent measurement on the board.

That is the whole tension worth reporting, and it is not an accusation. A vendor promoting a model to agent builders is normal. What is worth pinning down is the distance between the pitch and the numbers a reader can check, because the pitch is what a team acts on and the numbers are what the invoice and the wall clock deliver.

What the post actually establishes

Three things, and no more than three. The rest is inference dressed as news.

• The endorsement is real and dated. The post carries a 17 November 09:31 UTC timestamp, it comes from the vendor's own account, and it quote-tweets the platform's earlier announcement rather than standing alone. A vendor reposting a partner's launch is a deliberate act; the account did not simply happen to mention it.

• The platform is not a second deployment. Its model page lists qwen/qwen4-max in a table headed "Routed models served by their own labs" — not among its own published editions. That distinction matters more than the marketing copy around it: a routed model is Alibaba's deployment reached through someone else's front door, which is the only way a closed 2.4-trillion-parameter-class model can appear anywhere at all.

• It is a beta, not capacity. The site's standing banner says it is in private beta and asks for an invite code. An endorsement from the model's own account is a signal about intent, not about throughput, quota or a service-level agreement — and an invite code is not a procurement path.

And what the post does not establish is a single performance claim. There is no tokens-per-second figure in it, no task completion rate, no cost-per-task, no statement about how many concurrent agent sessions the route holds. Read it as "you may now reach this model here," which is all it says.

The card underneath the endorsement is cheaper than the vendor's own

The routed listing carries its own numbers, and the price is the one worth staring at. It reads $1.40 per million input tokens and $4.20 per million output, with a 977K-token context window and a status of online.

• Input — $1.40 on the routed card against $1.413 on Alibaba's own Global-scope card (Beijing, Hong Kong, Frankfurt, US Virginia, Tokyo) and $1.50 in Singapore.

• Output — $4.20 against $4.24 Global and $4.50 Singapore.

• The direction — the routed number sits roughly 1% below the vendor's cheapest published region and about 6.7% below the card most readers outside mainland China will actually be billed on.

• The stated mechanic — the platform describes its charge as "the lab's rate plus the route fee, a signed route receipt on every answer."

• The context figure — 977K on the listing, against 1,000,000 on Alibaba's model card, where the operative ceilings are 991,808 input tokens in standard mode and 983,616 in thinking mode.

Those two facts sit awkwardly together. A route that adds a fee on top of the lab's rate should cost more than going direct, and this one costs less than every published Alibaba region. The plausible explanations are unglamorous — a promotional rate, a regional arbitrage on which Alibaba endpoint is being called, or a listing that has not been updated since the preview card — and none of them can be confirmed from the outside. What a reader can do is stop treating the routed price as the model's price. Alibaba's own card is the reference; anything below it is a commercial arrangement with an expiry date nobody has published. If cost per task is the thing you are deciding on, size it off $1.50 in and $4.50 out in Singapore, or $1.413 and $4.24 on Global scope, and treat a discount as upside rather than as the plan.

A screenshot of the Alibaba Cloud Model Studio documentation page for qwen4-max, captured 17 November 2026 in English, showing the model identifier, the 1,000,000-token context window, a 991,808-token maximum input with a 983,616-token thinking-mode figure, a 131,072-token maximum output with 262,144 in thinking mode, a 131,072 default thinking budget, and per-region pricing of $1.5 and $4.5 per million tokens in Singapore against $1.413 and $4.24 in the Global scope.

The workload in the post is the one the tracker measures worst

Artificial Analysis is the only outfit publishing independent measurements of Qwen4-Max, and its page was updated with a full performance row that was not there on launch day. What that row says is not ambiguous.

• Intelligence Index — 43 on the v5.0.1 nine-evaluation suite, ranked 38th of 81 models, marginally above the tracker's median of 41 for the price tier.

• Output speed — 54 tokens per second. That is 74th of 79 models the tracker ranks on performance.

• Time to first answer token — 179.79 seconds, because the model thinks before it speaks and the default reasoning effort on Alibaba's card is high.

• Time per Intelligence Index task — 32.2 minutes, against a median of 13.3 minutes for comparable models. The tracker calls this "notably slow."

• Verbosity — 127 million output tokens across the index, against a median of 78.6 million. That is 65th of 80.

• Cost per index task — $3.27, ranked 44th of 80, with a 95% cache discount already applied.

Every figure in that list is the tracker's own measurement, not Alibaba's. The vendor's published claims are a different set entirely and are labelled as vendor claims wherever they appear: a 24% to 35% token-efficiency gain over Qwen3.8-Max on its own Work360 benchmark, 48% to 49% on Terminal-Bench 4.0, and average cost per task falling from $4.43 to $1.67. None of that has been independently reproduced. The two datasets do not contradict each other — a model can be efficient per task and still be slow per token — but only one of them tells you what an agent loop will feel like at 2am.

There is a second gap in the vendor's own material, and it is a strange one for a model being sold on agentic coding. Alibaba's model card does not publish a parameter count; the launch material describes a 2.4-trillion-parameter sparse mixture-of-experts with 71 billion activated per token, and the tracker records the size as undisclosed. There is also no published Terminal-Bench 4.0 score for this model from Alibaba itself, only the relative efficiency claim against its predecessor. A vendor that wants agent builders to adopt a model usually leads with an absolute agentic number. This one led with a ratio.

What 32 minutes per task and 54 tokens a second do to a loop

Unit prices hide this, so here is the arithmetic on the shape the post is advertising — a coding agent that works through a task in turns. All of it is derived from the published rates and the measured speed, not from a bill.

• Generation time per turn — a 3,000-token patch and reasoning block at 54 tokens per second is about 56 seconds of pure output. At 8,000 tokens it is roughly 2.5 minutes, before the model has read anything.

• A 25-turn session — 25 turns of 3,000 output tokens is 75,000 tokens, about 23 minutes of generation at that speed, plus 25 separate first-answer-token waits of nearly three minutes each if the harness does not keep the context warm. The overhead alone can exceed the generation.

• The same session in money — 75,000 output tokens at $4.50 per million is $0.34. The compute is cheap; the clock is not. That inversion is the single most useful thing to understand about this model's economics.

• Cache is the lever that decides whether it is tolerable — a 200,000-token context re-read cold at $1.50 per million costs $0.30 per turn; cached at the 95% discount it costs about $0.015. Over 25 turns that is the difference between $7.50 and $0.38 of input, which dwarfs the output bill entirely.

Put differently: on Qwen4-Max the thing to instrument is not output length, it is cache hit rate and first-token latency. A loop that keeps its context warm and tolerates a three-minute think is workable. A loop that re-sends a cold 200K context every turn pays twenty times the input rate and waits three minutes each time to do it. The 32.2-minute index task is the tracker's composite, not your task, but the direction of it is the thing to plan around.

A screenshot of the Artificial Analysis model page for Qwen4 Max, captured 17 November 2026 in English, showing an Intelligence Index of 43 at rank 38 of 81, a performance rank of 74 of 79, time to first answer token of 179.79 seconds, output speed of 54 tokens per second, 32.2 minutes per index task, 127 million output tokens, a $3.27 cost per index task, and a 984k context window.

The receipt does not say what answered

This is the part of the routed listing that deserves more attention than its price. The platform's own documentation is unusually candid: every routed answer carries a signed route receipt, and the receipt "does not prove which weights answered." It records where a request was sent and what the most recent probe reported. Models the platform publishes itself get a stronger chain — files stored by SHA-256 digest, revisions identified by a file-map hash, signed records under a pinned publisher key, an append-only transparency log — and the verification status shown for qwen/qwen4-max is "unknown."

For a closed model this is not a scandal, it is a structural limit that every router shares, ours included: if the weights are not published, no third party can prove which weights produced a token. What follows is a practical rule rather than a criticism. If your work depends on the model actually being Qwen4-Max — an eval, a benchmark submission, a contractual claim about which model touched the data — a routed endpoint is not sufficient evidence on its own, no matter how good the receipt looks. If what you need is the capability, a routed endpoint is fine and the provenance question is academic.

How to try it without betting a production path on it

Being straight about where we sit: OrcaRouter does not route Qwen4-Max. The catalogue carries qwen/qwen3.8-max, anthropic/claude-fable-5.2 and openai/gpt-6.1-sol; a request for a qwen4 identifier returns not-found today, so if you want this model on our key you are waiting like everyone else. What we can do is make the evaluation cheap to reverse.

The reason to care is that everything about this launch argues against hardcoding it. The model is a day old, its only independent performance row ranks it 74th of 79 on speed, its routed price undercuts its own vendor's card for reasons nobody has explained, and the platform carrying the endorsement is in private beta behind an invite code. Any one of those can change inside a week — and a preview-to-GA transition that also cut the shared thinking-plus-answer budget from 393,216 tokens to 262,144 and moved billing from credits to per-token metering is exactly the kind of change that breaks a production path in week two.

That is the case for keeping the experiment behind a routed endpoint rather than a hardcoded one: the new model on one route, qwen/qwen3.8-max on the fallback, one key and one bill for both. With 0% markup and the provider's list price passed through, the arithmetic you do against Alibaba's card is the arithmetic you get billed — which also means a vendor price cut shows up on our side the same day rather than at the next contract renewal. Model fusion is the other option worth knowing about for an unproven model: run it alongside the incumbent on a sample of your own traffic and compare outputs before either one owns the path. If you would rather just see what one API for 200+ models looks like, the catalogue is public.

A generated two-column scoreboard titled 'Qwen4-Max — the pitch and the meter', with rows comparing what the launch material claims against what Artificial Analysis measured: agentic coding endorsed on 17 November against a performance rank of 74 of 79, a 1,000,000-token context against 984k recorded by the tracker, a vendor token-efficiency ratio against an Intelligence Index of 43 at rank 38 of 81, no published speed figure against 54 tokens per second, no published task time against 32.2 minutes per index task, and average cost per task cut from $4.43 to $1.67 against $3.27 per task measured by the tracker. Footer reads 'Vendor figures unreproduced; speed, cost and index figures per Artificial Analysis, 17 November 2026.'

What would change this read

Three things, in order of how much they would matter.

The first is a speed figure from anyone other than the one tracker. A single 54 tokens per second measurement on a composite benchmark is a thin basis for a production decision, and the vendor has not published one at all. If a second independent harness lands near that number, the slowness is a property of the model and agent builders should size their loops around it. If it lands far above, the tracker's endpoint was the problem and this section dates badly — which is the right way for it to age.

The second is the verification status on that routed listing moving off "unknown," or the platform leaving private beta. Either would turn an endorsed route into something you could put behind a support commitment.

The third is Alibaba publishing an absolute agentic coding score rather than a ratio against its predecessor. Until it does, the endorsement is a claim about intent, and the only published measurement of the workload it advertises is 32.2 minutes a task at 54 tokens a second. That is not a reason to avoid Qwen4-Max — it is a reason to benchmark it on your own tasks before you promise anyone a delivery date, and to keep the incumbent one route away while you do.