· via dev.to (home feed)
Alibaba's Qwen4-Max undercuts rival pricing but drops open weights for its flagship
Alibaba's Qwen4-Max pairs a 2.4-trillion-parameter design with API pricing well below Anthropic and Google, but the flagship's weights stay closed. Key figures come from Alibaba itself and await outside verification.
Alibaba has released Qwen4-Max, the first model of the Qwen4 generation and the largest the company has built so far. According to a dev.to report, the model was unveiled on November 15 and became generally available on November 16. What stands out is less the scale than two deliberate choices: pricing well below Western competitors and, for the first time on a Qwen flagship, closed weights.
A sparse giant with a long context window
The model's backbone contains 2.4 trillion parameters, but only 71 billion are active for any given token. This mixture-of-experts arrangement routes each request through just the relevant subset of the network rather than engaging the whole thing, which keeps per-query compute, and the resulting bill, far below what the headline parameter count implies.
Qwen4-Max also handles images and video natively and accepts up to one million tokens in a single prompt, enough to load an entire codebase at once, as the dev.to post notes.
Pricing aimed at agent workloads
Through the QwenCloud API, Qwen4-Max costs $1.50 per million input tokens and $4.50 per million output tokens. The dev.to comparison puts Anthropic's Claude Opus 5.5 at $4 and $20 for input and output respectively, and Google's Gemini 4 Carbon at $2 and $10, with Google's lower rate scheduled to double on December 1.
Alibaba additionally claims the model needs up to roughly 50 percent fewer tokens than its predecessor, Qwen3.8-Max, to complete the same work. On the company's own Work360 evaluation, average cost per task fell from $4.43 to $1.67, a 62 percent cut. The dev.to post flags that Work360 is Alibaba's benchmark and that nobody outside the company has verified the figure.
Security treated as a training objective
According to Reuters, cited in the dev.to piece, Alibaba ran security checks alongside coding tests during training, targeting code injection, memory-safety flaws and broken access controls. The company says it filtered training data by purpose, keeping material useful for defensive work such as vulnerability discovery while removing data that could sharpen offensive capability, and that an internal review found no sign of large-scale cybersecurity risks.
On a test of whether generated code is both functional and secure, the pass rate reportedly rose from 22 percent to 31.8 percent. That figure is also Alibaba's own measurement, and it still leaves most attempts failing.
Closed weights, unverified claims
The biggest break with Qwen tradition is access. Earlier Qwen models could be downloaded and self-hosted; Qwen4-Max cannot. It is reachable only through Alibaba's cloud and its apps, namely Qwen Studio, QwenWork and Qoder. Smaller Qwen4 models are promised with open weights eventually, but no date has been given.
The second caveat is provenance. Every notable number in the launch, from benchmarks to cost savings to safety results, comes from Alibaba itself, with no independent laboratory or external red team involved so far. That leaves the claims unconfirmed rather than disproved.
Why it matters
The launch captures two shifts at once. First, the frontier-model contest is increasingly about cost per task rather than raw intelligence alone, a metric that matters enormously for coding agents and other pipelines making hundreds of API calls a day; Qwen4-Max's low token prices plus claimed token efficiency aim squarely at that economics. Second, Alibaba, the vendor most identified with open-weight frontier models, has gated its flagship behind its own cloud and metered access. If outside testing validates even part of Alibaba's numbers, a cheaper default for heavy workloads could emerge; if it does not, the release will mainly illustrate how much of today's pricing narrative rests on vendor-reported figures.
- #alibaba
- #qwen4-max
- #large-language-models
- #api-pricing
- #open-weights