Lilith.
⌕
Editorial illustration: Qwen4 Max accelerated the chase, but open weights still trail the frontier
Lilith illustration · editorial remix

Alibaba launched Qwen4 Max, which scored 43 on the Artificial Analysis Intelligence Index. That is a 9-point improvement over Qwen3.8 Max, but it does not change the open-weight story. Qwen4 Max is available through an API.

Qwen4 Max gained 9 points but still trails Kimi K3.1 Pro

Artificial Analysis tested Qwen4 Max at xhigh reasoning effort. It scored 43 on the Intelligence Index, compared with 34 for Qwen3.8 Max. Kimi K3.1 Pro remains ahead at 46.

The same evaluation puts Qwen4 Max's cost at $3.27 per index task, roughly 11% below Kimi K3.1 Pro at $3.68. Qwen4 Max has a context window of about 1 million tokens, accepts text and image input and is available through Alibaba's first-party APIs. These are results from one independent benchmark suite, not a universal model ranking.

A cheaper API does not solve the open-weight plateau

Ethan Mollick pointed to a different part of the story. In his view, more than 8 months after the Fable and Mythos class was first made public, no open model has reached that level. Qwen4 Max does not alter the argument because Alibaba has not released its weights.

The distinction is practical. An API can deliver strong performance and a low price without giving customers the ability to run the model on their own infrastructure, modify its weights or control long-term vendor dependence. Progress in inference pricing should not be confused with progress in access to frontier capabilities.

One leaderboard cannot measure control, privacy or operational risk

Artificial Analysis combines tasks covering knowledge work, coding, science and general knowledge. A score of 43 is a useful comparison inside that methodology, but it does not answer questions about data sovereignty, auditability or API stability.

Fable class is not a standardized technical category either. Mollick uses it to describe a capability tier, not a precise threshold that one score can certify. His claim is a plausible diagnosis of the market, but it remains an interpretation over a fast-changing set of models and benchmarks.

Open-weight releases must prove themselves in both work and operations

The meaningful shift will come when someone releases weights for a model that approaches the closed frontier across several independent evals and remains practical to operate. The license, required GPUs, inference cost and performance inside real workflows will all matter.

For now, Qwen4 Max demonstrates rapid capability gains and better API economics. It does not answer when that level will become available as open weights.

Lilith's verdict

Qwen4 Max made the front-row ticket cheaper, but the weights remain behind the curtain. A cheaper seat still does not give the audience the keys to the theater.