Qwen 3.7 Max — Alibaba's agentic flagship with a 1M context
Alibaba unveiled Qwen 3.7 Max on 20 May 2026 at the Alibaba Cloud Summit — the most expensive and most capable model in the Qwen family so far, positioned explicitly as an agentic flagship. It ships with a one-million-token context window, a native extended-thinking mode, and an API price well below the Western frontier models. Unlike earlier Qwen releases, Qwen 3.7 Max is closed-weight: there are no open weights, and the model runs only through the API. It is available via Alibaba Cloud Model Studio (DashScope), OpenRouter and Together AI.
What Qwen 3.7 Max actually brings
- 1M-token context and a native extended-thinking mode that is on by default for Max and can be tuned per request.
- Closed-weight, API-only — the first time a major Qwen flagship ships without open weights. Access via Model Studio, OpenRouter and Together AI.
- Price: $2.50 / 1M input tokens and $7.50 / 1M output tokens (OpenRouter, codersera) — exactly half the Opus 4.7 rate on input.
- Autonomy demo: 35 hours running autonomously, 1,158 tool calls with no human intervention, roughly a 10x speedup on a GPU kernel the model had never seen during training (vendor-reported).
What used to be the case
Alibaba’s Qwen models were known above all as an open-weight alternative: strong Chinese models you could download, self-host and fine-tune. That openness was a key selling point against the closed systems from OpenAI, Anthropic and Google. If you ran a Qwen model, you kept control over weights and deployment.
At the same time, the leading agentic models — the ones that work through tasks across many tool calls — were long seen as the domain of Western vendors. Claude Opus and GPT-5.5 set the bar on long-running coding and agent tasks, and Chinese labs trailed behind on most leaderboards. Qwen 3.7 Max takes aim at exactly these two points.
What’s the case now
With this release, Alibaba shifts its own positioning in three ways:
1. The flagship is closed for the first time. Qwen 3.7 Max ships without open weights and is usable only through the API. On its top model, Alibaba now follows the Western pattern: full performance is available as a service, not as a self-hosted download. The open Qwen line still exists below it, but the strongest model is now a product, not a download.
2. The price clearly undercuts the Western frontier. Qwen 3.7 Max costs $2.50 per 1M input tokens and $7.50 per 1M output tokens (OpenRouter). For comparison, Claude Opus 4.7 sits at $5 / $25 per 1M tokens (CloudZero). On input, Qwen is exactly half the price; on output, only about a third. A temporary 50% launch discount briefly cut the rates to $1.25 / $3.75; cached input is normally $0.25 per 1M tokens. For long, agentic runs with a lot of repeated context, that cache discount adds up on top.
3. The agentic benchmarks aim straight at the top. According to the vendor and public leaderboards, Qwen 3.7 Max scores 60.6 on SWE-Pro, 69.7 on Terminal-Bench 2.0 and 92.4 on GPQA Diamond — figures said to land ahead of DeepSeek V4 Pro and Claude Opus 4.6 on agentic coding. On the Artificial Analysis Intelligence Index it sits at 56.6, making it the strongest Chinese model in that ranking. In the LMArena evaluation, the Max preview ranks 13th in text (Elo around 1,475) and 7th in math.
Analysis
What is genuinely new about Qwen 3.7 Max is not a single benchmark, but the kind of task it is marketed for. Alibaba demonstrates the model with a 35-hour, fully autonomous run: 1,158 tool calls with no human intervention, ending in a roughly tenfold speedup on a GPU kernel the model had never seen during training. That is a different class from “answer a question” — here the model is shown as an autonomous worker that plans, executes, checks and corrects over hours. A one-million-token context window exists precisely for this: an agent running for hours has to carry its own history, intermediate results and tool outputs.
This framing calls for caution. The most spectacular numbers — the 35-hour run, many of the agent benchmarks — come from Alibaba’s own measurements, and independent reproduction is, per the reviews, only just beginning. Vendors pick demo tasks that make their model look good; an internally measured tool-call chain says little about how reliably the model holds up over hours on an unfamiliar, awkward task. The leaderboard scores (SWE-Pro, Terminal-Bench, LMArena) are more checkable because they sit on public rankings — but they too measure narrowly defined tasks, not the day-to-day of an agent working in a real codebase. What is solid above all is the price: $2.50 / $7.50 is a hard, verifiable argument, regardless of how the model ultimately performs.
What you can do now
If you build long-running agents: Test Qwen 3.7 Max on your own representative tasks against your current model — not against the vendor’s demo. A 35-hour demo is no substitute for a test on your codebase. Measure success rate, token usage and cost per completed task, not just the list price.
If your cost block is output-driven: This is where the gap is largest. Output at $7.50 instead of $25 per 1M tokens can matter a lot on output-heavy workloads (long generations, many iterations) — provided the quality is enough for your use case.
If you depend on open weights: Qwen 3.7 Max is not an option for you — it is closed-weight and API-only. Stay with the open Qwen models below it or with other open-weight frontier models and compare there.
Compare models systematically
How to weigh models against each other by context, price, license and fit is laid out here: → Comparing AI models
Entdecke mehr
Usage-based instead of flat-rate: The AI cost turn was predictable — and here is how to absorb it
AI providers are pushing flat rates toward metered billing. Why it had to happen, what it costs and three levers that soften the shift.
Glossar$/MTok (Cost per Million Tokens)
Standard pricing unit for AI APIs — cost in US dollars per one million processed tokens. Listed separately for input, output and sometimes cache.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.