Update: GLM-5.3 gets a per-token price — $1.40/$4.40 on OpenRouter
Z.ai's frontier coding model was coding-plan-only at launch. Now it's metered with 1M context — but thinking is forced on and the weights stay locked.
Copy markdown
The price, by the token
GLM-5.3 now lists at $1.40 in / $4.40 out per million tokens, with cached reads at $0.26 — the per-token rate Z.ai withheld at its Aug 14 launch. It went live Aug 18 as a single-provider (Z.ai) passthrough on OpenRouter.
1M context without the subscription
The metered endpoint exposes the full 1,048,576-token window and 131,072-token max output, so you can run big-context agent jobs pay-as-you-go instead of committing to the $18–$160/month GLM Coding Plan.
Thinking is mandatory now
Reasoning is always on and can't be turned off — calls that send the old `thinking: disabled` flag will fail. Port from GLM-5.2 by setting `reasoning_effort: low` for the cheapest, fastest mode.
The catch: throughput
OpenRouter clocks the hosted route at ~24 tokens/sec with a 6.3s median first-token latency — far below the 115 tok/s Z.ai advertises on its direct API. Fine for background agents, sluggish for anything interactive.
Why the open weights are still locked
GLM-5.3's cyber capability isn't hypothetical: it reportedly surfaced a 'serious vulnerability' in Cursor during testing. Z.ai is staging API access and holding the open weights until a safety review clears — roughly two weeks after the Aug 14 launch.
Elsewhere: x-ray any Hugging Face model
A Show HN tool, modelmap.cc, renders an interactive, animated architecture diagram for any Hugging Face model — a quick way to eyeball attention layout and size before you commit to fine-tuning one.