Alibaba's Qwen4-Max Opens to Pay-As-You-Go API, Task Costs Down Over 60%
Alibaba's Qwen4-Max is now callable from the standard pay-as-you-go API, lifting a preview restriction that had kept it on fixed-price plans. The company says the average cost per task has dropped by more than 60% against the previous model, betting that price, not leaderboard rank, decides which AI gets used.
- Original source: note.com.
- The source published it on Nov 15, 2026 at 8:13 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Qwen4-Max is available immediately on QwenCloud and can also be used in Qwen Studio, QwenWork and Qoder, according to posts from Alibaba's official Qwen team on X. Pricing is set at $1.50 per million input tokens and $4.50 per million output tokens. Cached input costs $0.075, which Alibaba says is 70% cheaper than before. A preview landed on November 11 but was limited to fixed-price plans and could not be called from the standard pay-as-you-go API — that lock is what just came off.
The model runs on a new MoE architecture and handles images directly, with a context window of 1 million tokens. Alibaba's post puts the model at 2.4 trillion parameters. Weights for the largest version will not be released, and smaller versions are promised for later.
On Alibaba's own numbers, Qwen4-Max is level with or slightly ahead of top North American models in the coding metrics the company presented. It scored 75.7 on DeepSWE, where Opus 5.5 scored 74.2, and 46.7 on Agents' Last Exam against 43.4 for the other figure. Bloomberg reported it outperformed Claude Fable 5.1 and Kimi K3.1 Pro in coding. Those results come from Alibaba's internal evaluations and have not been independently verified by third parties.
Alibaba also says the win isn't only unit price: the model burns fewer tokens to finish the same job, producing the average per-task cost cut of over 60%. Cached input is the lever — tasks that attach the same preamble or output format on every request get that repeated chunk at the lower rate, while long, differently worded instructions each time barely benefit. Reaction on X was muted, with prominent voices saying they would wait for third-party verification and that output felt unchanged from the preview stage.
TopicsAlibaba · Qwen4-Max · QwenCloud · Bloomberg · Opus 5.5 · Claude Fable 5.1 · Kimi K3.1 Pro · Qwen Studio
Source: note.com
Written by AI from verified reporting.

