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

OrcaRouter on the AA 43 Bill: 105K Tokens a Task

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

43 is the score; $3.27 is the bill. On November 16, Artificial Analysis admitted Qwen4-Max to its Intelligence Index (v5.0.1, xhigh effort, 43 points) and reported a weighted average cost of $3.27 per index task. OrcaRouter reverse-engineered that bill into tokens the same day: how much text one task burns, where the money goes, and how much of the official "62% cheaper per task" comes from the price cut versus from saying less.

104,940 output tokens per task

OrcaRouter's reconstruction puts one index task at 104,940 output tokens on average — 68,914 of reasoning and 36,025 of final answer. All three figures are roughly 30–40% lower than the previous generation: total output down 39%, reasoning down 42%, answers down 30%. The new model is not only cheaper; it talks less, and the convergence is sharpest in the reasoning portion. Note that these token counts are OrcaRouter's estimates (the author states the assumption that cache writes are billed at full price), not raw counts published by AA — read them with that caveat attached.

Cache writes overtake cache reads as the biggest cost

The cost mix is more interesting than the total: cache writes account for 51% of a task's cost, overtaking cache reads at 34%, and becoming the single largest line. That means Qwen4-Max generates a lot of new cache content across AA's nine evaluations instead of repeatedly hitting old cache — contexts keep refreshing in long-horizon tasks, and the write bill follows. For users, negotiating the input price is not enough: the cache-write price shapes the real bill too.

The 67% cut: about half price, about a third discipline

The official per-task cost fell from $4.43 for the previous generation to $1.67, a drop of about 62% (AA's $3.27 per-task figure is a different number, because AA's task set and weights differ — the two should not be compared directly). OrcaRouter splits the saving: about half comes from the price cut itself and about a third from token discipline — fewer outputs, tighter reasoning. The remainder comes from differences in task mix and accounting. In other words, Qwen4-Max's cheapness is not only a pricing decision; part of it is the model talking less.

Still verbose: 127M against 78.6M

But thrift is not the same as frugality. Across the full index run, AA counts 127 million output tokens from Qwen4-Max, against a 78.6 million median for models in the same price tier — about 60% more. The roughly 11% lower cost per task ($3.27 against Kimi K3.1 Pro's $3.68) comes from the lower unit price, not from fewer words. The honest reading of 43 against 46 is a slightly lower index score at a slightly lower task cost.

Two side notes

First, Qwen3.8-Max has been marked deprecated by AA — the previous generation has formally left the index, so the 43-versus-34 comparison (on the v5.0.1 scale) is now a historical figure. Second, AA has moved to re-running the index every 18 hours, so 43 is not a one-off snapshot but a floating reading; the score and the cost can both move on the next run, and any citation should carry the version and the date.

Our take: 43 on its own is only "above the 41-point median for its price tier". What is worth remembering is the bill's structure: 105K output tokens per task, cache writes over half the cost, and about a third of the saving coming from token discipline. At $1.50/$4.50 it is the cheaper pick in its tier — but if your tasks also keep generating fresh context the way AA's evaluations do, the cache-write price will rise first. Measure it 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 Token Efficiency Math (Nov 16, 2026) and Artificial Analysis — Qwen4 Max model page (AA Index 43, v5.0.1)

← Back to the Qwen4-Max topic page