GLM-5.2: a solid million-token context

GLM-5.2 was listed on OpenRouter on 16 June 2026 with a context window of 1,048,576 tokens and a maximum output of 943,718 tokens. Z.ai calls it a substantial leap in long-horizon capability over GLM-5.1 and says that, for the first time in the family, the capability holds on a "solid 1M-token context" that stably sustains long work.

A million tokens is roughly a large codebase, several long reports or hours of transcripts in one request. The practical benefit is that an agent can keep the whole project in view instead of summarising and forgetting.

Long context is expensive to serve, so GLM-5.2 changes the architecture. Z.ai proposes IndexShare, which reuses the same indexer across every four sparse-attention layers; the model card says this reduces per-token FLOPs by 2.9× at a one-million-token context length.

GLM-5.2 also improves its multi-token prediction (MTP) layer for speculative decoding, increasing the acceptance length by up to 20% according to Z.ai. Higher acceptance means more tokens are confirmed per step, which speeds up generation.

GLM-5.2 adds multiple thinking effort levels so you can balance quality against latency. A quick lookup does not need the same reasoning budget as a tricky refactor. On OpenRouter the model accepts the reasoning and reasoning_effort parameters.

Prices on OpenRouter are unusual: about $0.034 per million input tokens and $7 per million output tokens. Reading huge inputs is cheap, while long answers cost more — worth keeping in mind for agent workloads that produce a lot of output.