GLM-5.1: built to keep working for hours

GLM-5.1 was listed on OpenRouter on 7 April 2026. Its description makes a specific point: earlier models were built around minute-level interactions, while GLM-5.1 is meant to work independently and continuously on a task. The Hugging Face card puts it this way — the model "sustains optimization over hundreds of rounds and thousands of tool calls".

In practice that means the kind of work an engineer hands to an agent: break a problem down, run an experiment, read the result, find what blocks progress and try again. Z.ai says GLM-5.1 handles ambiguous problems with better judgment and stays productive over longer sessions than GLM-5.

On OpenRouter, GLM-5.1 has a context window of about 205K tokens and can return up to about 131K tokens per request. It accepts text and returns text. Pricing is $1.40 per million input tokens and $4.40 per million output tokens, a little higher than the original GLM-5.

The weights are published on Hugging Face as zai-org/GLM-5.1 under the MIT licence. Z.ai does not list a separate parameter count for GLM-5.1 on the card, so we do not quote one here.

GLM-5.1 was the long-horizon specialist of spring 2026. Since then GLM-5.2 has extended the context to about one million tokens and GLM-5.3 has improved coding further on the same base as GLM-5.2. For most new work, those newer models are the better default.

GLM-5.1 still makes sense if you already run it in an agent stack and your prompts are tuned for it, or if you want an MIT-licensed long-horizon model with a smaller context footprint. As always, test it on your real workload before switching.