Z.ai’s GLM-5.3 targets longer coding jobs, with a notable API change

Z.ai says GLM-5.3 improves coding-agent performance and long-running task handling, while using fewer output tokens on its in-house coding benchmark. For developers upgrading, there is a catch worth spotting before anything breaks: reasoning can no longer be switched off.

Zhipu AI / Z.ai GLM Watch analysis

What happened

Z.ai’s GLM-5.3 documentation describes the model as available to GLM Coding Plan users. The company says it is built on the same base model as GLM-5.2, with improvements from post-training, and supports text input, a one-million-token context window and up to 128,000 output tokens.

The documentation gives users three reasoning-effort settings: low, high and max. Applications that previously sent thinking.type: "disabled" need to switch reasoning on and set the effort level, or requests will fail. Z.ai recommends low as the replacement setting for those applications.

Our top picks

• Stronger long-horizon coding scores
  Z.ai reports Terminal-Bench 3.0 rising from 4.6 to 28.3, and DeepSWE v1.1 from 46.2 to 66.9, versus GLM-5.2.
• More work for fewer output tokens
  On Z.ai’s Code Bench, GLM-5.3 at max effort scored 34.5% using about 75,000 output tokens per task; GLM-5.2 scored 23.4% at about 96,000.
• Multiple reasoning effort levels
  Developers can choose low, high or max effort, though reasoning itself must remain enabled.
• A very large context window
  The documented one-million-token context and 128,000-token output limit could suit projects that need substantial material in one request.
• Off-peak Coding Plan discounts
  Z.ai says calls during off-peak hours, including weekends, use 50% of the standard points.

Why it matters

The useful promise is not simply a higher score: Z.ai’s figures suggest GLM-5.3 can complete more demanding coding tasks while using fewer output tokens on its own benchmark. That could matter for developers choosing a model for agent workflows, where a task’s total usage can count more than a headline price per token.

The API change is more immediate for existing users. Teams that rely on disabling reasoning will need to update their requests before changing the model ID. And the performance numbers are Z.ai’s benchmark results, not a guarantee that every codebase or workload will see the same gains.

Our read

There is enough here to make GLM-5.3 a serious candidate for a developer’s own task-based evaluation, rather than a model to adopt on the strength of a leaderboard alone. Test it on representative work, compare both success and usage, and check the reasoning setting before upgrading. A model can be clever; an avoidable failed request is still an avoidable failed request.

What to watch

• How GLM-5.3 performs on independent evaluations and real-world coding tasks.
• Whether Z.ai adds support for disabling reasoning in a later iteration.
• Whether lower output-token use translates into lower total costs for agent workflows.

Discussion spark: When assessing coding models, should teams prioritise task success, token use, or the ease of fitting the model into their existing tools?

not affiliated with or endorsed by Zhipu AI, Z.ai or GLM