Send GLM-5.3 a screenshot and it has nothing to look at. Z.ai's current flagship takes text, and only text. That is an unusual choice in a year when most new models read images by default, and it tells you what the model is for: code and agent work.
Z.ai's release notes date GLM-5.3 to August 18 and a second model, GLM-5.3-Flash, to August 26. This entry is late. It was added during the October catch-up, with every figure checked against the live documentation on the verification date above.
The API model is glm-5.3. It has a 1M-token context and up to 128K output tokens. It always reasons, at low, high, or max effort, and Z.ai says reasoning can no longer be switched off. Plan for that on small edits, because every request pays for some thinking. Model documentation.
The list price is $1.40 input / $4.40 output per million tokens, with cached input at $0.26. That output rate is less than half of Sonnet 5.5's $10. Whether a finished task is cheaper depends on how many reasoning tokens the model spends getting there. Z.ai pricing.
Z.ai reports a 50% gain over GLM-5.2 on its own Code Bench, Terminal-Bench 3.0 rising from 4.6 to 28.3, and DeepSWE v1.1 rising from 46.2 to 66.9. It also reports 84.5% on CyberGym and a doubled ExploitBench score. These are vendor results, and one of the benchmarks is the vendor's own. Use them as a reason to run a trial.
The weights are on Hugging Face: a 753B-parameter mixture-of-experts model under a license the card calls the GLM-5.3 license. Read that license and size the hardware before treating self-hosting as a real option. The hosted API is the simpler way to find out whether the model suits your work.
If you need images, look at GLM-5.3-Flash instead. It accepts video, images, files, and text, keeps the 1M context, and lists at $0.15 / $0.50 per million tokens. Z.ai describes a 320B-parameter model with 18B active. A faster FlashX variant costs $0.37 / $1.25. Flash documentation.
GLM-5.3 is now in the model tools, marked text-only and “not tested.” Flash is not in the calculators yet. A fair first trial is a text-only bug with a failing test: fix the effort level, record the output tokens, and compare the patch with the one your current model writes.