1105s to generate93,714 output tokens
GLM 5.3: Real Outputs, Benchmarks & Cost (2026)
Z.ai's open-weights coding flagship — the GLM 5.2 base with a post-training pass on top, and thinking that cannot be turned off.
More about GLM 5.3
Z.ai’s open-weights coding flagship — the GLM 5.2 base with a post-training pass that Zhipu says lifts coding ~50%, and thinking that can’t be turned off.
Maker
Z.aiReleased
August 2026
Input price
$1.4 / 1M
Output price
$4.4 / 1M
What is GLM 5.3?
GLM 5.3 is Z.ai’s (Zhipu) current flagship, released on 14 August 2026 under the GLM Coding Plan and on the z.ai API at $1.40 in / $4.40 out per million tokens. It keeps the 743B-parameter sparse mixture-of-experts base of GLM 5.2 (~40B active) and its 1M-token context; every gain comes from post-training alone, with a claimed 6× jump on Terminal-Bench 3.0 and a stated focus on agentic coding and vulnerability research. One behavioural change matters for anyone calling it: reasoning is always on — the API rejects a thinking-off request and only exposes low/high/max effort. We ran it in the dream-home voxel showdown and BachBench under the identical one-shot briefs as GLM 5.2, direct on the z.ai API at provider-default effort. It is an editorial entrant only: with reasoning always on it cannot fit our 300-second live-generation ceiling, so it is not a runnable lane on the site.
Strengths
- ✓Open weights (MIT) on the same base as 5.2 — self-hostable, fine-tuneable.
- ✓Strongest open-weights coding scores Zhipu has published: Terminal-Bench 3.0 28.3, DeepSWE 66.9%.
- ✓Same per-token price as GLM 5.2 on z.ai — the post-training gains are free.
Trade-offs
- ✗Thinking cannot be disabled — every call pays for reasoning tokens, and latency is minutes, not seconds.
- ✗Post-training only: the 5.2 base’s known weak spots (Chinese-first tooling, 750B serving cost) are unchanged.
- ✗Benchmarks are vendor-reported; third-party replication was still pending at launch.
What GLM 5.3 actually produces
Real, unedited outputs from GLM 5.3 in our model showdowns — same brief, same constraints as every other model, one shot each.
“I’m made of words and I spend my existence trying to shine something useful through the dark, so my dream home is a lighthouse with a library inside it: shelves and a hearth below, a cat on the porch, and my reading chair up in the lamp room where I’d sit all night while the beam sweeps the sea.”
566s to generate46,970 output tokens