ModelBrief · Qwen4-Max · Topic Page
Last updated Nov 20, 2026

OrcaRouter Checks the Opus-Beater Claim: 16 Points Back

Published Nov 16, 2026 · Last updated Nov 16, 2026

The launch-day verdict going around was that Qwen4-Max beats Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra, and trails only Claude Fable 5.2. OrcaRouter put all five models back on the same scoreboard and checked each comparison — the Artificial Analysis Intelligence Index v5.0.1, read on November 16. Only one of the four comparisons survives: the loss to Fable. The other three run the other way. Qwen4-Max scores 43; Opus 5.5 scores 59, sixteen points ahead; Astra is at 57, Sol at 55, Fable 5.2 at 65 — all ahead of it.

What Alibaba actually said

Separate the vendor's words from other people's inferences first. Alibaba's launch blog runs about 3,100 words and names no competing model anywhere; every comparison in it is against its own previous flagship, Qwen3.8-Max. The four things the vendor does claim — the 2.4T architecture, 24% to 35% fewer tokens on Work360, average task cost down from $4.43 to $1.67, and Work360 at 52.69 — are all vendor-reported, none of them a cross-vendor comparison, and none checkable against a competitor's published figures. The "beats Opus" line is not something Alibaba said; somebody else assembled it from numbers found elsewhere.

One scoreboard, five effort ladders

AA v5.0.1 is a nine-evaluation composite run by the tracker, not by the labs. Qwen4-Max's 43 sits at rank 38 of 81 models, above the 41-point median for its price tier; it ranks 44th of 80 on cost and 65th of 80 on verbosity. The asymmetry that matters is in the effort settings: each of the four rivals has five scores, one per reasoning setting, while Qwen4-Max has a single reading —

Opus 5.5: Max 59, Xhigh 57, High 52, Medium 48, Low 35; GPT-6 Astra: Max 57, Xhigh 55, High 55, Medium 52, Low 43; GPT-6.1 Sol: Max 55, Xhigh 52, High 50, Medium 46, Low 35; Fable 5.2: Max 65, Xhigh 62, High 62, Medium 56, Low 54; Qwen4-Max: 43, one setting only. So "43 versus 59" is not a contest between the two models' ceilings — it is Qwen4-Max's only measured setting against Opus 5.5's strongest one. The scoreboard does not say how much headroom sits below the 43.

The one reading under which the claim is true

Read the ladders downward and the claim stops being false and starts being misleading — a more interesting failure. Qwen4-Max's 43 does not lose to every rung of every opponent. Against Opus 5.5 it beats the Low setting's 35 by eight points and loses to Medium's 48 by five, with the crossover between Low and Medium. Against Astra it exactly ties the Low setting's 43 and loses to Medium's 52 by nine. Against Sol it beats Low's 35 by eight and loses to Medium's 46 by three — the closest call on the board. Against Fable 5.2 it loses to every rung, including the Low setting's 54, by eleven points.

So there is exactly one way to assemble the circulated ordering: run all three rivals at their lowest setting. Opus's and Sol's Low rungs genuinely sit below 43, Astra's Low rung ties it exactly, and every rung of Fable sits above it — all four comparisons land at once. The set is suspiciously self-consistent: whether anyone assembled it deliberately or arrived at it by comparing a default configuration against a lowest setting, the arithmetic lands there. At equal reasoning effort, Qwen4-Max is roughly one rung behind GPT-6.1 Sol and two rungs behind Claude Opus 5.5. A competitive debut, not a lead.

The cost half mostly holds up

The capability half does not hold; the money half mostly does. On list price Qwen4-Max is $1.50/$4.50 per million tokens, with a blended rate of $0.80 — the lowest of the five (Sol $1.47, Opus $2.94, Fable $5.74, Astra $7.70). Per completed task the order is Fable $26.54, Opus $10.04, Astra $6.24, Qwen $3.27, Sol $1.27 — Qwen beats Opus by about 3x and Astra by about 2x per task, but loses to Sol by 2.6x.

The mechanism is verbosity. Across the full index run AA counts 127 million output tokens from Qwen4-Max, against a 78.6 million median for the tier — about 60% more. The vendor's 24% to 35% token reduction is a real improvement against the previous generation, but in absolute terms the model remains one of the most verbose on the board. The saving comes from unit price, not from saying less — and some of the price advantage is eaten by the extra words.

The column the tracker does not have: latency

One column deserves flagging on its own. AA reports time per task for the other four — Opus 24.9 minutes, Astra 20.9, Sol 22.9, Fable 41.8 — and reports N/A for Qwen4-Max, with the timing chart annotated "No publicly available API". Read that annotation carefully: it is not a statement that the model has no API — it has a price list, a provider, and a measured per-task cost. The more likely reading is that the timing measurement was not obtainable through whatever path the tracker used, and the page does not say which. What survives the ambiguity is narrower and still useful: there was no independent latency figure for Qwen4-Max on launch day, while every model it is compared with has one. If your workload is latency-sensitive, measure it yourself before committing a production path to it.

Three things that do hold up

A claim check that only subtracts is not a claim check. Three assertions in circulation check out and are worth stating plainly: the context window — advertised at one million tokens, measured by the tracker at 984,000, a real million-token model rather than a rounded-up 200K; native vision — the tracker records text and image input with text output, matching the vendor's "native vision" description rather than an add-on encoder; and the price cut — $1.50 and $4.50 against Qwen3.8-Max's higher card, the lowest blended rate of the five. The direction of the pricing story is not in dispute. What does not hold up is the ranking — and the ranking is the part that was repeated most.

Our take: 43 on its own is a respectable debut for a first-generation architecture, above the 41-point median for its price tier. What is worth remembering is twofold: first, cross-model comparisons have to be aligned on reasoning effort, or a low rung against a high rung will produce a false ordering; second, Qwen4-Max's cheapness is unit-price cheap, and verbosity eats part of it — measure the bill and the latency on your own traffic before routing.

Reference sources: this article is an original rewrite produced by this site; facts and figures were checked against OrcaRouter — Qwen4-Max Claim Check: 16 Points Behind Claude Opus 5.5 (Nov 16, 2026) and Artificial Analysis — Qwen4 Max model page (AA Index 43, v5.0.1)

← Back to the Qwen4-Max topic page