Alibaba's Qwen3.7-Plus Beats GPT-5.4 at Screen Control for One-Sixth the Price

Most frontier AI labs have kept vision and coding separate: you pick a vision model for image tasks, a coding model for agents, and you stitch them together yourself. Qwen3.7-Plus breaks that pattern. Alibaba's Qwen team released Qwen3.7-Plus, a multimodal agent model that combines visual perception, graphical user interface control, and code generation within a single autonomous agent loop. The result is a model that can look at your screen, decide what to click, write the code to act on it, run that code, check the output, and loop until the task is done , all without switching models or pipelines.

One model, two worlds

Unlike most frontier models that separate vision and agent capabilities into distinct offerings, Qwen3.7-Plus is designed to perceive, reason, code, and act across both GUI and CLI environments simultaneously. That distinction matters in practice. Alibaba built Qwen3.7-Plus on top of the Qwen3.7-Max foundation and added what the language-only Max doesn't have: eyes. The model can read your screen, understand what's on it, navigate graphical interfaces, automate browsers, and operate desktop applications from screenshots.

It can understand visual interfaces, perceive on-screen content, and perform both GUI interactions and CLI operations, while also leveraging environmental feedback for code generation, application manipulation, testing, validation, and iterative optimization. By integrating the full workflow of "see, think, write, act, and verify" into a unified agent loop, it enables end-to-end automation of complex software tasks from initial understanding to final delivery.

The agentic loop is powered by five concrete abilities on top of image and video understanding:

• Deep reasoning , multi-step planning before acting
• Self-programming , the model writes and revises its own code
• Tool invocation , it calls external functions or APIs
• Verification and testing , it runs outputs and checks results
• Autonomous iteration , it loops until the task is done

The benchmark that actually matters

ScreenSpot Pro measures GUI grounding , the model's ability to look at a screenshot and pinpoint exactly which pixels to click on. It is the bottleneck capability for everything we call "computer use" in 2026. State-of-the-art scores sit in the 75–82 range; Qwen3.7-Plus at 79.0 puts it in the frontier tier alongside Claude Computer Use and OpenAI Operator.

Here's how the key numbers stack up against the competition:

OmniDocBench 91.4% and OCR-Bench-V2 English 70.7% , above GPT-5.4 at 59.1% and Gemini-3.1 Pro at 64.6% , mean Qwen3.7-Plus processes real-world business documents like invoices, contracts, research papers, and financial reports with best-in-class accuracy.

Under the hood

It uses early fusion training on trillions of multimodal tokens, meaning vision and language are processed together from the first layer rather than integrated after the fact. This is architecturally different from bolting a vision encoder onto a text model , the model doesn't translate images into text descriptions and then reason; it reasons directly over the combined visual and linguistic signal from the start.

A hybrid architecture combining Gated Delta Networks with sparse Mixture-of-Experts keeps inference costs manageable across long agent loops. Gated Delta Networks are a linear attention variant , instead of the standard transformer's quadratic cost (where processing doubles as context grows), Gated DeltaNets improve focus sequence processing so the model can process super long text efficiently without drifting off or forgetting what matters, allowing it to efficiently process extremely long sequences with memory and computation scaling almost linearly with sequence length. That matters enormously for agents that need to sustain multi-hour autonomous runs.

Reinforcement learning was scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability. In plain terms: the model wasn't just trained on static examples. It was trained by running millions of agent simulations and learning from what worked.

The 11-hour demo that explains the ambition

Alibaba's internal demonstration of Qwen3.7-Plus's Hybrid-Agent system maps directly to the benchmark scores. The task: build a complete English vocabulary learning app, from scratch, end to end. The agent operated continuously for over 11 hours without human intervention. It generated more than 10,000 lines of code, triggered over 1,000 agent calls, and completed every stage of a real software development lifecycle.

Built on Qwen3.7-Plus, a browser agent is demonstrated through Qwen for Chrome, a browser extension embedded in Chrome. Users can interact with Qwen directly from the browser sidebar and, with authorization, switch it into Agent mode. In this mode, Qwen can perceive the current webpage, understand the user's task, plan the next steps, and operate as a browser agent to perform clicks, typing, navigation, configuration, and verification directly in the real browser environment.

Where it fits (and where it doesn't)

The practical use cases break down clearly:

• GUI automation , browser agents, desktop app control, mobile app navigation end-to-end
• Document intelligence , invoices, contracts, research papers with search-augmented QA
• Agentic coding , full dev cycles from requirement analysis to testing and iteration
• Visual reasoning , math from diagrams, chart analysis, screenshot-to-code
• Productivity workflows , multimodal input across text, images, and video

Use Qwen3.7-Max if you're tuning for pure-text SWE-Bench Pro scores or need the fastest text-only latency. Max handles text reasoning and code execution; Plus handles screen perception and visual workflows. They are designed as complementary endpoints, not competitors.

The price changes the math

Qwen3.7-Plus bolts vision and video understanding onto the Qwen 3.7 text backbone, runs a one-million-token context window, and lists at roughly one-sixth the per-token price of the text-only Qwen 3.7 Max. Concretely: $0.40 per million input tokens, $1.60 per million output tokens, and $0.08 for cached input , roughly six times cheaper than Qwen3.7-Max.

The pricing is roughly 5 to 10x cheaper per million tokens than OpenAI for comparable capability. For teams building agent pipelines that make thousands of tool calls per session, that gap is not cosmetic , it's the difference between a prototype and a production system.

One meaningful caveat: the model is proprietary and API-only; no open weights shipped with it. Alibaba has historically released open-weight versions of its Qwen models weeks to months after the flagship API launch. An open-weight Qwen3.7 variant is likely but not yet confirmed.

What this means for the field

Qwen3.7-Plus perceives screens, reads interfaces, writes code from visual references, calls tools, runs tests, checks its own output, and iterates until the task is complete , across both GUI and CLI environments, within a single context window, without switching models or pipelines. That is a fundamentally different architecture than vision-augmented text generation, and it changes what you can actually build.

The model is now generally available through Alibaba Cloud's Bailian platform (marketed internationally as Model Studio) via API. It is also accessible through OpenRouter. The assumption that computer-use agents require closed, expensive frontier models is now harder to defend , Alibaba has put frontier-tier GUI grounding at budget-tier pricing, and that changes the calculus for anyone building agents in 2026.