GLM-5V Turbo: Z.ai's Vision Model for Screenshots and GUI Agents

Most vision-language models are evaluated on describing images: captions, chart questions, OCR. That is a different job from looking at a screenshot of a broken layout and writing the CSS that fixes it, or clicking through an Android app to complete a task. GLM-5V Turbo is aimed squarely at the second kind of work.

Z.ai calls it its first multimodal coding foundation model, built for vision-based coding and agent workflows in tools such as Claude Code and OpenClaw. The positioning is clear. The published detail behind it is thinner, and that shapes how you should evaluate it.

What it takes in, and what it is for

According to Z.ai's model page, GLM-5V Turbo accepts video, images, text and files, and outputs text only. Context is 200K tokens with up to 128K output, and it supports thinking mode, function calling, streaming and context caching. It first appeared on OpenRouter as z-ai/glm-5v-turbo at the start of April 2026; on Z.ai's API the id is glm-5v-turbo.

The stated use cases cluster around three jobs:

• Design to code: turning a mock-up or screenshot into working front-end code.
• GUI agents: understanding a screen, planning an action, executing it, then looking again.
• Visual exploration: browsing websites on its own and turning what it finds into code, or debugging a rendering problem from a screenshot.

Z.ai also ships official skills for it on ClawHub, covering image captioning, visual grounding, document-grounded writing, resume screening and prompt generation. Those are a reasonable hint at where Z.ai expects it to be used.

What is disclosed about the model, and what is not

The architecture notes are qualitative. Z.ai describes native multimodal fusion using a new vision encoder called CogViT, a multi-token prediction architecture for inference efficiency, and reinforcement learning optimised jointly across more than thirty task types, from STEM and grounding to video, GUI agents and coding agents. It also says agentic capabilities were introduced during pre-training rather than only in fine-tuning.

What it does not give is a parameter count. The page says it reaches strong results with a smaller model size than comparable models, but puts no figure on either. There is also no published weights repository on Z.ai's Hugging Face organisation and no licence statement, so treat it as an API-only model. If you need to self-host a vision model, look elsewhere in the GLM line.

If you want background on how screenshots become tokens a language model can attend to, multimodal models explained covers it. The short version: images and especially video consume input tokens quickly, so the per-token price understates the cost of a screen-heavy agent.

Benchmarks you will have to verify yourself

Z.ai's page names the evaluations it reports: design-to-code generation, visual code generation, multimodal retrieval and question answering, AndroidWorld and WebVoyager for GUI and web agents, CC-Bench-V2 across backend, frontend and repository exploration, plus PinchBench, ClawEval and ZClawBench for agent work. The results are presented as charts rather than numbers in text, so we are not quoting scores here.

The list itself is informative. AndroidWorld and WebVoyager measure whether an agent can complete real tasks in a phone or browser, not whether it can describe what it sees. CC-Bench-V2 and the three agent benchmarks are text-only, so they show whether the vision work held up on plain coding. That is the right set of tests for the stated job, but like any vendor-run suite it describes a best case.

Picking between GLM-5V Turbo and its neighbours

OpenRouter lists GLM-5V Turbo at $1.20 per million input tokens and $4.00 output at the time of writing, identical to the text-only GLM-5 Turbo. That makes the sibling choice simple to state:

• Pick GLM-5V Turbo when the agent needs to see: screenshots in a UI-fix loop, design mocks, GUI automation, video.
• Pick GLM-5 Turbo when the workload is text-only OpenClaw tool chains. Z.ai tuned it specifically for that, so do not assume the vision model matches it on pure text without running both.
• Look at cheaper vision options when the images are routine. Z.ai's pricing page lists GLM-4.6V at $0.30 and $0.90, and GLM-5.3 Flash, which Z.ai's release notes credit with native visual capabilities, at $0.15 and $0.50. For captioning and document reading the price difference is large. The GLM-5.3 Flash guide covers that model.

Vision-language models compared has a broader checklist for choosing across vendors.

A design-to-code test you can run in an afternoon

Because the benchmark numbers are not in text, your own test carries more weight than usual. A practical one:

1. Collect eight screenshots of real components from your product, with the source for each. Include two with dense tables and one with a modal over content.
2. Ask the model to reproduce each from the screenshot alone, in your actual stack and design tokens.
3. Render the output and compare it visually against the original. Note layout errors separately from styling errors.
4. Then run a fix loop: give it a screenshot of a deliberately broken state plus the source, and ask for a diff. Count how many rounds it needs.
5. Log input tokens per request. Screenshot-heavy loops are where the bill surprises people.

If it gets layout right and needs one or two fix rounds, it is earning its rate. If a cheaper vision model gets you to the same place with an extra round, the cheaper model probably wins on cost per finished component.