Consider three workloads on the same paid API. One pastes the same 80-page contract into every request. One lets an agent run a long chain of tool calls against a single repository. One fires off short, unrelated questions all day.
The first two get cheaper. The third does not. That is not a billing bug, and it is the design bet behind Alibaba's newest flagship.
What actually shipped
On 16 November 2026, Alibaba released Qwen4-Max, the top model in its Qwen4 series. The company's published specifications put total parameters at 2.4 trillion, with native vision and a context window of one million tokens. It is live on Alibaba's Qwen platform as a paid API, and it is also reachable through the group's Token Plan and inside its office, coding and consumer apps.
Two numbers in that release need separating, because most coverage blurs them.
- 2.4 trillion is the size of the parameter pool the model can draw on.
- 71 billion is what gets activated for any single token — roughly 3% of the pool.
The gap between them is what Alibaba is selling.
The 3% trick, and why it is not free
Qwen4-Max is a mixture-of-experts model built on what Alibaba calls Qwen Sparse Attention. Instead of running every parameter for every token, a routing layer selects a small subset of expert modules. Alibaba says each token activates about 25% fewer parameters than it did in Qwen3.8-Max.
The trade-off is familiar to anyone running these systems in production. Sparse routing saves compute per token, but the full parameter pool still has to be resident somewhere: memory footprint does not shrink in proportion to activation. Batch efficiency depends on whether traffic arriving at the same moment happens to route to the same experts. A sparse model that looks cheap on paper can be expensive on a badly shaped workload.
Alibaba's own efficiency framing points at a different metric than list price: tokens burned to finish a job. On Terminal-Bench 4.0, a professional coding evaluation, the company reports that average token use per attempt fell from about 289,000 to about 151,000 at the same "xhigh" thinking setting, while the score reached 48.5. Fewer tokens per success compounds, because it hits compute, latency and price at once.
The number Alibaba is selling, and the one that matters
Cached input is priced down to RMB 0.5 per million tokens (≈ US$0.07 / HK$0.55). Alibaba puts the drop in cost per completed task, versus Qwen3.8-Max, at 58%, which is the figure behind its "roughly 60%" framing. Caching is one of the mechanisms that would produce a gap that size; the company presents the total, not a breakdown.
According to Li An, chief scientist at BrainNet (脑机网), the number that matters in this release is not the parameter count but the tokens spent per finished task, because that is the figure an enterprise budget actually tracks.
Caching rewards a specific shape of work: long, stable prefixes that repeat across requests — a fixed system prompt, a codebase context, a document a team keeps asking about. It does little for a stream of one-off, short, unrelated prompts. So the headline discount is real, but it is not a uniform price cut across the product. It is aimed at the workloads enterprises actually run: the same policy document, the same repository, the same customer record, asked about a thousand times.
Where it sits
Alibaba reports Qwen4-Max placing sixth on CodeArena, a blind front-end coding comparison, and cites agentic and professional-task benchmarks — including Agents' Last Exam, where it reports 46.7 — as evidence the model sits in the leading group. These are company-selected evaluations with company-reported scores. They are directionally useful and not a substitute for running your own workload.
The open-weight question is where some coverage has read Alibaba's plan as covering the flagship. It does not. Alibaba's stated commitment is that the small and medium-sized models in the Qwen4 series will be released as open weights in stages. That is a promise about the smaller members of the family, not about the 2.4-trillion-parameter Max.
One market detail matters for anyone judging the release. API prices for Chinese domestic models were cut repeatedly through 2026, and promotional token packages were sold at steep discounts during the year's shopping festivals. A 58% cost-per-task improvement lands in a market where buyers already expect prices to fall, so the durable advantage is less about this month's number and more about whether efficiency gains keep arriving.
What this means outside China
First, the frontier is being redefined around cost per completed task rather than cost per token. A per-token price without a tokens-per-task figure describes half of the economics.
Second, sparse architectures make benchmark comparisons less portable. Two models with the same score can differ by multiples in what it costs to reach that score, depending on how prompts are shaped.
Third, the open-weight pipeline for smaller siblings is one of the main channels through which Chinese model design reaches developers outside China. Which sizes get released, and under what licence, is a better signal of competitive intent than a launch keynote.
Honest limitations
This article relies on Alibaba's published launch specifications and on Chinese business-media reporting of the release (Shanghai Securities News, The Paper, Beijing News' Shell Finance, Southern Metropolis Daily, QbitAI). All benchmark results, including the 58% cost-per-task reduction and the 48.5 Terminal-Bench score, are company-reported and were not independently reproduced here; outlets differ slightly on the previous generation's score, so this article cites only the new one. No independent third-party evaluation of Qwen4-Max was available at the time of writing, and no pricing outside the Chinese platform was confirmed. The claim that caching mainly benefits stable, repetitive workloads is analysis, not a company statement. Open-weight timing for the smaller Qwen4 models has been described only as "in stages," with no dates. The article does not cover safety evaluations, training-data provenance, or performance on non-Chinese, non-English tasks.
What readers can do now
- Measure tokens per finished task, not price per million tokens. Run one representative workflow end to end on your current model and count what it burns. That baseline is the only thing a vendor discount can be compared against.
- Audit how much of your traffic is cacheable. If prompts share long stable prefixes, a cached-input discount is worth real money. If every request is short and unique, headline discounts will not show up on the invoice.
- Wait for independent benchmarks before migrating a critical path. Company-reported evaluation results are a starting point for a trial, not a reason to move production traffic.
