Skip to content
NewsModel release

Z.AI's GLM-5.3 reaches OpenRouter with stronger coding and leaner token use than GLM-5.2

· by Pondero Newsdesk

The short version

Z.AI's GLM-5.3 became available via API on August 18, delivering 50% stronger coding scores than GLM-5.2 while using fewer output tokens per task. Open-source weights are expected around August 28.

Z.AI's GLM-5.3 reaches OpenRouter with stronger coding and leaner token use than GLM-5.2

Z.AI released GLM-5.3 via its developer API on August 18, a pure post-training upgrade over GLM-5.2 that scores 50% higher on coding benchmarks while generating fewer output tokens per completed task than either its predecessor or competing closed models at the same performance level.

What the model delivers

GLM-5.3 uses the same base model as GLM-5.2, which launched in June 2026 with IndexShare, a sparse-attention optimization that cuts per-token compute 2.9x at the full 1M-token context length, per InfoWorld's June 2026 coverage of that release. GLM-5.3 carries that same stack forward. Every performance gain over GLM-5.2 came from scaling post-training on longer, more complex coding environments, per Z.AI's August 14 blog.

The gains are concentrated in hard, multi-step coding tasks. On Terminal Bench 3.0, a benchmark for terminal-based software engineering, GLM-5.3 scored 28.3, up from 4.6 for GLM-5.2. On DeepSWE v1.1, it improved from 46.2 to 66.9. On Z.ai's in-house Code Bench, it reached 34.5% at Max reasoning effort, up from 23.4% for GLM-5.2, per the Z.AI blog.

Reasoning is always on and cannot be disabled. Three effort levels (low, high, max) control the reasoning budget. Z.AI recommends max for complex coding tasks, per the model documentation.

The token-efficiency story

GLM-5.3 reached those stronger results while generating fewer output tokens per task than comparable models at equivalent performance. At Max effort, it completed Z.ai Code Bench tasks using roughly 75,000 output tokens, compared with GLM-5.2's 96,000 at the same effort level, per the Z.AI blog. At High effort, it scored 31.4% using around 50,000 output tokens, surpassing Claude Opus 4.8's 29.5% on the same benchmark at 120,000 tokens.

For teams running long-horizon agentic coding loops, output token count is the variable that controls per-run cost when a model's price is fixed. A model that solves more tasks with shorter reasoning chains cuts that cost without requiring any infrastructure change.

How to access it now

GLM-5.3 went live on OpenRouter on August 18 from Z.AI's own endpoint, priced at $1.40 per million input tokens and $4.40 per million output tokens. It supports OpenAI's Chat Completion protocol, OpenAI's Response protocol, and Anthropic's Message protocol, per the GLM-5.3 docs, so teams already configured for other providers can swap in the model ID without rebuilding their scaffold.

Open-source weights were not published at launch. Z.AI said weights would ship approximately two weeks after the August 14 blog post, once safety evaluation and hardening are complete, per the Z.AI blog. That window points to around August 28.

Why it matters

The cybersecurity benchmark results (GLM-5.3 scored 84.5% on CyberGym, edging Mythos 5 and GPT-5.6 Sol) drew the most coverage at the August 14 announcement. The practical operator story is different: a model that improves task completion rates while generating shorter output sequences.

Teams that routed code through GLM-5.2 for long-context agentic tasks get a drop-in upgrade. Teams built on Claude SDK patterns can add GLM-5.3 as an alternate model without protocol changes because of the Anthropic Message protocol endpoint. Self-hosters waiting on the August 28 weights release will get a benchmark-verified checkpoint that can run on private infrastructure, relevant for compliance-constrained shops that cannot send production code to a cloud API.

What to watch next

The open-weights release around August 28 is the first checkpoint: independent researchers can then verify the Terminal Bench 3.0 and DeepSWE scores without relying on Z.AI's own runs. If additional OpenRouter providers take on the model, the routing and failover options will broaden from the current single-provider setup.

Sources