← Back to feed

Alibaba’s Qwen 3.8 family is now fully available through OpenRouter, but the four models within it exist in radically different market conditions. A September 11 snapshot from OpenRouter captures the split clearly.

Four Models, Four Market Structures

Max is the flagship and remains a closed, Alibaba-only offering on OpenRouter — one provider, fixed price. Flash has attracted two providers: Alibaba and Makora. 2.4T, the mixture-of-experts model, has seven providers, and six of those charge an identical $2.00 input / $6.00 output per million tokens. Venice is the outlier at $2.50 / $7.50.

27B is where the real market action is. Fourteen providers now serve it, and prompt pricing on September 11 ranged from $0.15 to $0.45 per million tokens — a 3x spread on the same underlying weights. Completion prices ranged from $2.00 to $3.20. The model was benchmarked by Artificial Analysis at xhigh reasoning effort for its intelligence score; the specific snapshot date matters because reasoning effort affects the measured output.

What the Spread Signals

A 3x prompt-price spread across 14 providers is a commoditization signal. When that many vendors compete on identical weights, margin compression follows. The 27B model is dense (no sparse routing), which means inference infrastructure costs are more predictable than MoE variants — easier for smaller providers to offer competitively.

The 2.4T MoE shows the opposite dynamic: seven providers, but six immediately converged on $2/$6. MoE models have variable compute costs depending on which experts activate per token, and providers either price cautiously or adopt the market rate set by the largest player (Alibaba) quickly.

The Closed-End Premium

Max sitting at a single provider with no competitive pressure is consistent with Alibaba keeping its highest-capability variant as a direct revenue line. At the other end, 27B’s democratization into a 14-provider commodity is a typical open-weight lifecycle: labs release weights, the ecosystem routes demand, and the price finds a floor within weeks.

For developers, the 27B selection now warrants a provider check before committing to an integration — a 3x difference on prompt tokens is material at scale. For the 2.4T, the market has already decided: $2/$6 is the clearing price until the next model generation displaces it.

Qwen 3.8 Max, 2.4T, and 27B scores from Artificial Analysis are based on snapshots; the Max row reflects a September 2 measurement (snapshot version qwen3.8-max-0902) and the 27B row was scored at xhigh reasoning effort. Flash was not included in Artificial Analysis’s benchmark comparison table for this family. Scores will move if Artificial Analysis remeasures or if a different effort setting is used.