GLM-5.3-Flash is out - a cheap open multimodal model built for agent work

The release

Z.ai (formerly Zhipu AI) released GLM-5.3-Flash on August 26. It is an open-weight model under the MIT license, which means you can download it, run it, and pin a version without asking anyone. It takes text, images, and video as input, holds a 1M-token context window, and uses a mixture-of-experts design with 18 billion active parameters out of 320 billion total.

Pricing on Z.ai's own API is $0.15 per million input tokens and $0.50 per million output, with a 50% launch promotion through September 9 that halves both. Artificial Analysis lists a blended rate near $0.10 per million. The model is also on OpenRouter and in Vercel's catalog, and it runs on vLLM and SGLang if you host it yourself.

What is actually new

This is the first GLM-5 model with vision built into the model rather than bolted on as a separate step. For document and screenshot work that matters. One model reads the PDF, looks at the embedded chart, and calls the next tool, instead of you wiring an OCR service to a vision model to a text model.

The agentic scores moved. Z.ai reports 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE, a software-engineering agent test, both up sharply from GLM-5.2. Independent testing from Artificial Analysis puts its Intelligence Index at 57, high for an open model at this size and price. On AutomationBench-AA, which scores whether an agent finishes real SaaS workflows without tripping a guardrail, GLM-5.3-Flash lands around 48.8. Competitive, though Google's Gemini 3.7 Flash still scores higher there at 62.7.

What it means if you run a business

A Flash-tier model is a worker. The job it fits is the high-volume, well-scoped task you already understand: pulling fields out of invoices, triaging inbound tickets, drafting first-pass replies, summarizing call transcripts, checking documents against a rule list. Work where you can define "correct" and measure it.

Two things make this release worth a look for that kind of work. The multimodal input removes a layer of plumbing from any pipeline that touches scanned documents or images. And the price, combined with open weights, changes the math on jobs that were too expensive to automate at frontier-model rates. At roughly $0.10 to $0.50 per million tokens you can run a pipeline thousands of times a day and still spend less than one salaried hour.

Open weights also buy protection from the thing this blog keeps writing about: hosted models get more expensive or get retired on the vendor's schedule. DeepSeek quadrupled a price three days after launch. Claude Sonnet 5 went up 50 percent. A model you host does neither.

The honest caveat

Most of the benchmark table is Z.ai's own testing. Treat it as a reason to run your own eval, not as a result.

"Flash" oversells the speed. Independent measurement puts output around 50 tokens per second, slower than average for its class, with a 1.5-second wait before the first token. Thinking cannot be turned off, so every call carries reasoning tokens you pay for and wait on. This is not the model for a live voice agent or anything a person watches load.

"Open" is not "free to run." The full model is 320 billion parameters. Self-hosting needs real GPU capacity, and for low volume the hosted API will beat the hardware bill. The ecosystem around this model is also new, so production reliability is unproven and tooling gaps are likely.

What to do about it

Pick one task you already run on a frontier model. Document extraction is the usual best candidate. Run a weekend eval against glm-5.3-flash through OpenRouter and compare two numbers: accuracy on 100 real examples, and total cost per run. If accuracy holds and the cost drop is large enough to matter, move that one task. Leave the rest of your stack alone until it earns the change.