Qwen 3.7 Plus: Alibaba's Low-Cost Multimodal Agent Model GA

Qwen 3.7 Plus is Alibaba's new low-cost multimodal agent model, generally available as of June 1, 2026. It 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 — a combination that positions it as a default for budget-sensitive agent pipelines.

The model first surfaced as Qwen3.7-Plus-Preview on the public LM Arena leaderboard around May 14, 2026, giving developers roughly 18 days of live inference signal before the commercial endpoint dropped the "-Preview" suffix at GA. That preview-then-commercialise cadence is becoming the standard release shape for Chinese AI labs, and it is part of what makes this launch worth a close read rather than a headline skim.

This guide covers what actually shipped, where the pricing sits in the multimodal budget tier, the GUI-grounding cost math that makes it interesting for automation teams, how to read its vendor-run benchmarks honestly, the independent signals that do exist, and the strategic pivot hiding in the "proprietary" label. Every benchmark table below carries an explicit vendor-stated caveat where it applies.

Qwen 3.7 Plus is the multimodal sibling of Qwen 3.7 Max. Where Max is a text-only flagship, Plus extends the same agentic backbone with vision and video understanding while keeping the coding, tool-use, and productivity strengths of the text model. Crucially, it is a perception model, not a generative one: it accepts text, images, and video as input and returns text only. It can read a screen, ground a click target, or answer a question about a frame of video — but it will not draw you a picture.

The model is exposed as the API endpoint qwen3.7-plus on Alibaba Cloud Model Studio (DashScope), reachable through OpenAI-compatible chat-completions and responses APIs across Beijing, Singapore, and US-Virginia regional endpoints. Alibaba's own framing positions it as a single agent that blends GUI and command-line interaction inside one loop. For the text-only counterpart, see our deep dive on the text-only Qwen 3.7 Max flagship.

The story of Qwen 3.7 Plus is economic, not heroic. At a list price of $0.40 per 1M input tokens and $1.60 per 1M output tokens, it costs roughly 6× less on input and 5× less on output than Qwen 3.7 Max ($2.50 / $7.50). VentureBeat positioned it as priced "just above" MiniMax M3's limited-time discount pricing, and the independent OpenRouter and Artificial Analysis listings corroborate the input/output figures. Cached-input pricing is the one number to treat with caution: OpenRouter and Artificial Analysis show $0.08/1M, while VentureBeat cites $0.04/1M from the Model Studio pricing page. Until the live page settles, model it as a $0.04–$0.08 range rather than a fixed rate.

For agent workloads, input price dominates total cost because tool loops, screenshots, and long context windows pour tokens through the prompt side far faster than the model emits them. A model that reads a full screenshot, a system prompt, and a running scratchpad on every step is paying input tokens dozens of times per task. Cutting that side from $2.50 to $0.40 per million is the difference between an agent that is economical to run at scale and one that is a demo. This is the same FinOps logic we lay out in our AI inference cost optimization playbook.

The single most interesting number Qwen reports is its ScreenSpot Pro score — a GUI-grounding benchmark that measures whether a model can name the exact pixel coordinates of a click target on a screenshot. Qwen claims 79.0, ahead of GPT-5.4 at 67.4 and Claude Opus 4.6 at 49.5 in its own table. That number is vendor-stated and was measured with thinking disabled, so it should anchor a hypothesis, not a purchase order. But it points at the real reason this model matters for robotic-process-automation and browser-automation teams: it pairs a high grounding score with a low price.

So we built a metric the vendor announcements and the aggregator sites do not publish. Take each model's ScreenSpot Pro score and divide it by its input price per million tokens — call it grounding points per dollar of input. It is a crude proxy, but it answers the operative question for automation teams: which model gives you the most click accuracy per dollar? On vendor-stated grounding scores against list input pricing, the gap is not close.

The launch post is dense with benchmark tables, and nearly all of them were produced by Qwen evaluating Qwen — frequently on internal harnesses where the lab is both the test subject and the operator. We reproduce a representative selection below, with the model leading on agentic and long-context retrieval tasks and conceding ground on pure-text software engineering. Orange bars mark Qwen 3.7 Plus where it leads its own comparison; blue bars mark where a frontier competitor leads.

The pattern is coherent. Qwen 3.7 Plus is strongest exactly where its positioning claims — agentic GUI and mobile tasks, terminal work, and long-context retrieval — and it gives up roughly 3 points of SWE-Bench Pro to its own Max sibling on pure-text software engineering. That is the expected trade for a cheaper, multimodal variant: you buy perception and price, you pay a little on text-only coding depth. On frontier-grade STEM the gap nearly closes, with a vendor-stated GPQA Diamond of 90.3 sitting within a point of Opus 4.6 Max and DeepSeek V4-Pro Max.

Strip away the vendor tables and only two independent signals remain, and both are worth more than the rest combined precisely because Qwen did not produce them. Artificial Analysis placed Qwen 3.7 Plus at #53 of 164 models on its Intelligence Index— its phrasing is "well above average" rather than frontier — with an output speed of about 52.9 tokens per second, which ranks roughly #101 of 164 and is notably slow, and a time-to-first-token near 2.3 seconds. The same evaluation flagged unusually high verbosity: the model emitted about 110M output tokens during testing against a 29M median, which matters because output tokens are the expensive side of the bill.

The second signal is LM Arena, where coverage placed the model around #15 on text, #12 on coding, and #16 on vision — enough for one outlet to rank Alibaba as the #5 lab in vision research at launch. Arena rankings fluctuate daily and these were reported in secondary coverage, so treat them as a snapshot rather than a standing. The honest synthesis: independent data confirms Qwen 3.7 Plus is a capable mid-pack model that is cheap and slightly slow — not the across-the-board leader the vendor tables imply.

Most coverage files "proprietary, API-only" as a footnote. We think it is the headline. Alibaba built the Qwen moat on permissively licensed open weights — US enterprises including Airbnb reportedly adopted earlier open-weight Qwen models, and that distribution was a strategic asset. Qwen 3.7 Plus ships with no open-weight checkpoints on Hugging Face, and Qwen 3.7 Max is closed as well. That is a deliberate pivot toward a closed-frontier posture, and it changes the calculus for any team that chose Qwen specifically because it could self-host.

Be precise about what is and is not known here. Alibaba has not confirmed open weights for the 3.7 family, and it has not denied them either. Reports of a Q3 2026 open-weight variant are third-party speculation, not an announced roadmap— do not plan around them. For the broader arc of the lab's release strategy, our retrospective on Alibaba's open-weight model history traces how a once open-by-default lab arrived here, and our framing of the wider closed vs open-weight AI trade-offs lays out what a closed Qwen costs sovereignty-bound buyers.

Three engineering choices make Qwen 3.7 Plus a genuine agent platform rather than a chat model with vision bolted on. First, the 1,000,000-token context window with up to 65,536 tokens of output and an internal chain-of-thought budget reported up to 256K tokens — enough headroom to hold a large codebase, a screenshot history, and a running plan in a single loop. Second, the preserve_thinking API parameter, which retains <think>blocks across conversation turns so a long-horizon task does not reset the model's reasoning chain on every tool call.

That last feature is quietly significant because it shows the whole frontier converging on the same idea. Anthropic ships Extended Thinking that returns reasoning blocks, OpenAI passes encrypted reasoning state back across turns, and Alibaba now exposes preserve_thinking. Multi-turn reasoning-state preservation has become table stakes for serious agent models — and Qwen 3.7 Plus offers it at the cheapest price of the three. Third, the model is positioned to generalize across agent scaffolds — Qwen cites Claude Code, OpenClaw, and Qwen Code among the frameworks it targets, though that cross-framework claim is vendor-stated and worth verifying against your own scaffold.

The model is not a universal answer. It is a sharp fit for a specific set of workloads and a poor one for others. Here is how we would route it across the four cases that come up most in client engagements.

Our standing recommendation is to treat Qwen 3.7 Plus as a routing target, not a default. Send screen-reading, GUI-grounding, and cheap-bulk long-context work its way; keep a stronger text coder for heavy software engineering; and reach for an open-weight model when sovereignty or self-hosting is non-negotiable. If you are deciding between budget multimodal options head-to-head, our coverage of MiniMax M3 is the natural companion read. And if you want this benchmarked against your own corpus and pipelines rather than a vendor table, our AI digital transformation engagements start with exactly that comparative eval.