Qwen 3.7 Max Pricing: $1.25/M on OpenRouter vs $1.66/M Direct

On May 20, 2026, Alibaba launched Qwen 3.7 Max — a model that immediately surged to #13 on the LMSys Chatbot Arena and racked up 721 points on HackerNews. It brings a massive 1-million-token context window, a new Agent Frontier designation, and a coding agent benchmark result that outperforms Claude Opus 4.6 Max and K2.6 Thinking. But what grabbed the developer community’s attention wasn’t just the performance — it was the Qwen 3.7 Max pricing. Within 24 hours, OpenRouter listed the model at $1.25 per million input tokens and $3.75 per million output tokens — a price that undercuts Alibaba’s own official API by a surprising margin. In this post, we break down every major way to buy Qwen 3.7 Max tokens, why the aggregator price is cheaper than direct, and where you should actually deploy it.

Qwen 3.7 Max Pricing Overview

Alibaba’s Qwen 3.7 Max represents the pinnacle of the Qwen family in mid-2026. Built with an Agent Frontier training recipe, it handles multi-step autonomous workflows, tool calling, and long-horizon reasoning without losing track of context. During internal testing, the model sustained a 35-hour autonomous coding session — a feat that demands both a robust 1M-token memory and reliable instruction following. These capabilities make Qwen 3.7 Max a prime candidate for coding agents, enterprise RAG pipelines, and AI assistants that need to digest entire codebases or legal documents in a single prompt.

However, frontier performance usually comes with a frontier price tag. The initial assumption was that Alibaba would charge a premium, especially given that the previous generation Qwen 3.6 Max already costs $1.04/$6.24 per million input/output tokens. The official Qwen 3.7 Max pricing on Alibaba Cloud’s 百炼 platform landed at 12元 per million input tokens and 36元 per million output tokens — approximately $1.66 and $4.97 USD. That’s a noticeable increase on the input side but a reduction on output compared to Qwen 3.6 Max. Still, the total cost of running complex agent workflows can add up quickly when you’re pushing 1M-token windows.

The surprise came on May 21 when OpenRouter added Qwen 3.7 Max at $1.25/$3.75. Suddenly, developers had a way to access the model for 25% less on input and 24% less on output than going straight to Alibaba. This unusual pricing inversion — where a reseller is cheaper than the source — has become the central discussion point around Qwen 3.7 Max pricing. In the following sections, we’ll dissect each provider’s offer, compare them against key competitors like deepseek-v4-vs-gpt-5-5-2026" style="color:#0284c7;text-decoration:underline">DeepSeek V4 Pro and glm-5-api-pricing-and-access-2026" style="color:#0284c7;text-decoration:underline">GLM-5.1, and explain the economics behind the numbers.

Alibaba Cloud (百炼) Direct Pricing

To access Qwen 3.7 Max directly from Alibaba, you need an Alibaba Cloud International account and access to the 百炼 (Bailian) platform. The official pricing is set in Chinese Yuan: 12元 per million input tokens and 36元 per million output tokens. Using a typical exchange rate of approximately 7.2 CNY per USD, this converts to $1.66 per 1M input tokens and $4.97 per 1M output tokens. Alibaba occasionally offers promotional credits, but the list price is what most enterprise customers will see.

Alibaba’s direct channel provides several advantages. You get dedicated throughput, enterprise-grade SLAs, data residency controls, and priority support. For large-scale deployments that need guaranteed latency and compliance certifications, the direct route is often mandatory. However, the signup process can be daunting for developers outside China: it may require business verification, Chinese-language documentation, and working with Alibaba’s regional sales team. The billing is also natively in CNY, which can introduce forex headaches for USD-based teams.

OpenRouter Pricing for Qwen 3.7 Max

OpenRouter made Qwen 3.7 Max available on May 21, 2026, just one day after the official release, at a rate of $1.25 per million input tokens and