GLM-5.3

GLM-5.3 is a GLM-5-series model released by Z.ai on August 14, 2026 for complex coding, long-horizon agent tasks, and cybersecurity work. Z.ai says it uses the same base model as GLM-5.2; the gains come from scaling post-training environments, task diversity, and reinforcement-learning compute over the preceding month.

Architecture and specifications

Because Z.ai explicitly identifies GLM-5.3 as using the GLM-5.2 base, DataLearner records the same approximately 753.33B-parameter MoE / DSA base. Z.ai has not separately disclosed the number of active parameters for GLM-5.3. The post-training stack carries forward IndexShare for long-context processing, SAO for reinforcement learning on long-horizon tasks, and the slime asynchronous training framework. Official evaluations use up to a 1M-token context and 128K maximum output on several long-running tasks, so DataLearner records a 1M context window and 128K maximum output. Input and output are text.

Thinking modes and API changes

The official model ID is glm-5.3. Thinking is always enabled: thinking.type=disabled is no longer supported. The reasoning_effort parameter accepts low, high, and max; the default is max, which Z.ai recommends for coding. Applications migrating from an older model must enable thinking before switching the model ID or the request will fail.

Official evaluations

Z.ai published a comparison table spanning coding, cyber, and agentic evaluations. GLM-5.3 scores 88.2 on Terminal-Bench 2.1, 28.3 on Terminal-Bench 3.0, 66.9 on DeepSWE v1.1, 58.0 on NL2Repo, 19.0 on ProgramBench, 78.1 on FrontierSWE, 42.5 on SWE-Marathon v1.1, and 39.8 on PostTrainBench. Cyber results include 84.5 on CyberGym, 105 and 130 solved ExploitGym instances under two-hour and six-hour budgets, and 54.4 on ExploitBench. Agentic results include 73.0 on Toolathlon Verified, 48.2 on AutomationBench v1.0.6, 28.5 on Agents' Last Exam, 62.5 on HLE with Tools, and 1769 on GDPval-AA v2.

These are vendor-reported release evaluations and may not reproduce across different deployments, sampling settings, or providers. Z.ai's footnotes show that most agent and cyber tests used Claude Code 2.1.207, max reasoning effort, and controlled tool environments, with benchmark-specific context limits, timeouts, and rollout counts. DataLearner maps these rows to the Max (With Tools) evaluation mode and preserves important conditions in the benchmark descriptions.

Cyber capability and planned weight release

Z.ai reports state-of-the-art vulnerability-discovery performance on CyberGym and particularly large gains over GLM-5.2 higher in the exploitation chain. Because these capabilities are dual-use, Z.ai plans to release weights about two weeks after launch, following additional safety evaluation and hardening. No verified GLM-5.3 weight repository or license is available yet, so DataLearner does not currently label the model as open source or inherit GLM-5.2's MIT license.

Access and pricing

GLM-5.3 is available to all GLM Coding Plan users and can be used through ZCode and coding agents such as Claude Code and OpenCode. Coding Plan uses credits: input, cached input, and output tokens have multipliers of 6.9, 1.7, and 24 per 10,000 tokens, with usage outside the weekday 14:00–18:00 UTC+8 peak window charged at 50% of standard credits. Z.ai's standard pay-as-you-go API pricing page did not list a per-million-token USD price for GLM-5.3 at launch, so DataLearner does not convert subscription credits into a currency price row.

Official sources

• GLM-5.3: Frontier Coding with Emergent Cyber Capabilities
• GLM Coding Plan overview
• Z.ai standard API pricing
• Official GLM-5.2 base model card