ModelsQwen3.7 Plus: The Mid-Tier Qwen Built for Screen-Reading Agents
Qwen3.7 Plus pairs a 1M window with vision aimed at GUI agents, at a fraction of Max pricing. What it claims, what it costs, and when an open 27B is the better call.
Most model releases describe vision as a feature: the model can look at images. Qwen3.7 Plus describes it as a job. Alibaba pitched it at agents that read a screen, decide what to tap or click, and carry on until a task in an app is finished. That is a narrower and more demanding claim, and it changes how you should evaluate the model.
It is also the 3.7 model that has aged best. Alibaba's Model Studio model list currently features Qwen3.7 Plus for text generation and visual understanding next to Qwen3.8 Max, which suggests the Plus tier was not refreshed in the 3.8 generation, at least not yet.
What it is
Qwen3.7 Plus launched in Model Studio on 1 June 2026, with the dated snapshot qwen3.7-plus-2026-05-26, and reached Alibaba's US region on 26 June. It is a hosted model with no published weights or parameter count.
Alibaba calls it a cost-effective Plus model with upgraded vision-language abilities that keeps the agent-level coding, tool use and productivity skills of the series. On OpenRouter it has:
- a 1M-token context window and up to 131,072 output tokens;
- text and image input, text output;
- a reasoning mode, tool calling, JSON mode and structured outputs.
Built around screens, not just images
The release note lists what Alibaba calls multimodal hybrid agent capability: perceiving real-world scenes, reading screens and operating graphical interfaces, generating code from visual references such as mock-ups, and navigating mobile apps end to end.
Those are four different tasks with different failure modes. Generating a component from a mock-up is one call with a checkable result. Operating a GUI is a loop: screenshot, decide, act, screenshot again, for dozens of steps. A model that is good at the first can still lose track in the second, so do not let a demo of one stand in for the other. Best model for vision tasks goes through what images are actually useful for in production.
Alibaba has not published benchmark numbers for these claims in the release notes we read. Measure them yourself.
The price shape of a GUI agent
OpenRouter lists Qwen3.7 Plus at $0.32 input, $1.28 output and $0.064 cache read per million tokens at the time of writing. Prompts of 256K tokens and above are listed at $0.96 input and $3.84 output.
GUI agents are input-heavy, because every step resends history plus a fresh screenshot. As an illustration, a 30-step task where each step sends a 20K-token prompt and returns 500 tokens costs roughly $0.0064 of input and $0.0006 of output per step, about $0.21 for the task. Cache the system prompt and tool schemas, which do not change between steps, and the input share falls.
The variable you cannot estimate in advance is how many tokens your screenshots cost at the resolution you send. Log the prompt token count per step on a real run before you trust any per-task figure, including this one.
Plus or Max
Qwen3.7 Max is listed at $1.475 input and $4.425 output, roughly 4.6 times and 3.5 times the Plus rates. On OpenRouter it is also text-only. For anything visual on that route, Plus is the 3.7 model you can actually use; the Qwen3.7 Max guide covers what the flagship is for.
Against its own predecessor, Plus is a straightforward upgrade on price: the same input rate as Qwen3.6 Plus, about a third less for output, and twice the output budget. The one thing 3.6 Plus has that 3.7 Plus does not list on OpenRouter is video input.
Below it sits Qwen3.7 Flash, much cheaper on short prompts. If your agent's steps fit under 32K tokens and the task is simple, try Flash first and escalate to Plus on failure.
Plus or an open 27B
The more interesting alternative arrived in August. Qwen3.8 27B is a dense vision-language model with weights under Apache 2.0, and Qwen reports 84.3 on OSWorld-Verified for it, a desktop-operation benchmark that is close to the job Plus is pitched at. That is a vendor figure on a vendor harness, but it is the kind of claim that should make you test both.
The trade is control against convenience. Plus is one endpoint at one price. The 27B is served by many providers at prices and quantisation levels that vary widely, or by you on your own hardware. If screen data cannot leave your network, the open model is the only one of the two that can meet that requirement. The Qwen3.8 27B guide covers the model and its provider spread.
How to test it for an agent loop
A short evaluation tells you more than any release note:
- Pick ten real tasks in the interface your agent will drive, from trivial to awkward.
- Run each three times on Plus with the same scaffold and screenshot resolution.
- Record success, steps taken, total prompt tokens and where it got stuck.
- Repeat on one alternative: Flash for cost, or the open 27B for control.
- Compare cost per completed task and, separately, the worst failure you saw.
For GUI agents the worst failure matters as much as the average, because a wrong click can do something you cannot undo. If Plus is consistent, it earns its place as the default. If it wanders after step fifteen, you have learned that before your users did.
Common questions
Can Qwen3.7 Plus operate a computer or phone interface?
Alibaba says so: its release note lists reading screens, operating graphical interfaces and end-to-end mobile app navigation. It has not published benchmark numbers for those claims in its notes, so test it on your own interface.
What does Qwen3.7 Plus cost?
OpenRouter lists $0.32 input and $1.28 output per million tokens at the time of writing, with $0.064 cache read. Prompts of 256K tokens and above are listed at $0.96 input and $3.84 output.
Is there an open-weight alternative to Qwen3.7 Plus?
Qwen3.8 27B is the closest: a dense vision-language model under Apache 2.0 that Qwen pitches at coding and agent work. It is a different model, so compare the two on your own tasks.
