Skip to content
CheapestInference

GLM-5.3 API — status, specs & subscription

GLM-5.3 is Z.ai’s (Zhipu AI) new coding and agentic model, released August 14, 2026 under the tagline “Built to Code. Ready for Cyber Defense.” It keeps the same 743B-parameter base as GLM 5.2 — every gain comes from scaled-up post-training — and Z.ai claims it as the strongest open-weights coding model it has measured, with a headline jump in cyber-security capability. The full analysis, the vendor benchmark table, and the head-to-head with GLM 5.2 are in GLM-5.3: specs, benchmarks, pricing & API options.

Quick facts

Section titled “Quick facts”
ModelGLM-5.3
Model IDglm-5.3
StatusReviewing — pipeline
Pool candidateFrontier Pool (upgrade path for GLM 5.2)
BaseSame 743B base as GLM 5.2 — gains from post-training only
Context windowZ.ai advertises a 1M-token variant (glm-5.3[1m], with compaction); standard-API spec not yet published
ReasoningEffort levels low / high / max (default max); thinking cannot be disabled
Headline claimStrongest open-weights coding model (vendor-run benchmarks); CyberGym 84.5
List price (reference)Not yet published — GLM 5.2 lists $1.40 in / $4.40 out per 1M
Open weightsPromised ~2 weeks after launch, pending Z.ai’s own safety evaluation — no weights, no license, no model card today

Why it’s a Frontier candidate

Section titled “Why it’s a Frontier candidate”

We already serve GLM 5.2 in the Frontier Pool, and GLM-5.3 is the same base model post-trained further — on Z.ai’s numbers, a big step in exactly the workloads the pool is bought for: agentic coding (Terminal-Bench 2.1 81.0 → 88.2, SWE-Marathon 19.4 → 42.5) and tool-driven automation (Toolathlon 59.9 → 73.0). If the review completes, the natural outcome is an in-place upgrade of the pool’s GLM slot — the same path DeepSeek V4 Flash took in the Core Pool with the 0731 build.

What’s gating it is availability, not quality signals: Z.ai has committed to open weights roughly two weeks after launch, once its own safety evaluation and hardening of the model’s cyber capabilities is complete — but today there are no weights, no license text, and no model card. GLM 5.2 shipped MIT; that precedent does not automatically set 5.3’s terms. Our own quality evaluation on real coding and agent workloads runs in parallel.

Follow the status

Section titled “Follow the status”