Z.ai GLM-5.3

Overview

GLM-5.3 is Z.ai's August 14, 2026 flagship for complex coding, long-horizon agent work, and vulnerability research. It uses the same base model as GLM-5.2; Z.ai attributes the gains to one month of expanded post-training across more environments, more varied professional tasks, and greater compute. That makes this an execution-focused upgrade rather than a new pretraining generation.

The headline result is a 50% improvement over GLM-5.2 on Z.ai's private Code Bench, alongside large gains on public terminal, repository, automation, and cybersecurity evaluations. GLM-5.3 is available to all GLM Coding Plan users and through ZCode at launch. Open weights are not available on launch day: Z.ai says they will follow after two weeks of safety evaluation and hardening.

What's New

Stronger Coding from Scaled Post-Training

Z.ai expanded training environments from coding exercises toward complete units of expert work. A task can include access to codebases, internal documentation, storage systems, compute clusters, and experiment output, requiring the model to diagnose a bottleneck, implement a change, run tests, and verify an end-to-end result. Research and judge agents help synthesize runnable tasks and check that their reward signals cannot be satisfied by shortcuts.

On Z.ai Code Bench, GLM-5.3 improves task completion and checklist accuracy at every effort level while using fewer output tokens. At Max effort, it reaches 34.5% at about 75K output tokens per task versus GLM-5.2's 23.4% at 96K. At High effort, it reaches 31.4% at about 50K tokens, compared with 29.5% at 120K for Claude Opus 4.8 in Z.ai's test.

Emergent Cybersecurity Capability

Vulnerability-discovery environments produced especially large gains further along the exploitation chain. Z.ai reports 84.5% on CyberGym, 54.4% on ExploitBench, and 105/130 completed ExploitGym tasks under normalized two-hour/six-hour budgets. The company also says expert-reviewed model runs found 2,436 vulnerabilities across 269 real-world projects, including 1,097 medium-to-high-severity issues, with disclosures tracked in its public security ledger.

These are defensive-research capabilities with meaningful dual-use risk. Organizations should restrict them to systems they own or are explicitly authorized to test, apply human review, and follow coordinated disclosure processes.

More Efficient Long-Horizon RL

GLM-5.3 continues using the IndexShare architecture, Single-Rollout Asynchronous Optimization, and slime training stack introduced around GLM-5.2. Z.ai added sampling controls, training–rollout consistency work, hierarchical local caching, dynamic teacher switching, and workload-aware scheduling. The company reports more than 2.3× higher end-to-end RL training throughput on long-horizon coding tasks.

Three Mandatory Thinking Effort Levels

GLM-5.3 supports low, high, and max reasoning effort, with max as the default and Z.ai's recommendation for coding. Thinking itself is always enabled. Applications that previously sent thinking.type: "disabled" must change configuration before switching model IDs or the request will fail.

Performance Benchmarks

The following results come from Z.ai's launch evaluation. They show the scale of the change from GLM-5.2 but remain sensitive to harnesses, context windows, timeouts, and vendor methodology.

GLM-5.3 does not lead every row in Z.ai's comparison table. Closed models remain ahead on several exploitation and coding evaluations, while GLM-5.3's strongest relative gains appear against its own predecessor.

Migration Guide

For GLM Coding Plan and coding-agent integrations available at launch, use glm-5.3 with thinking enabled. Z.ai currently marks the general GLM-5.3 API as coming soon, so this example should not imply launch-day general API availability:

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "max"
}

Before upgrading:

1. Search client configuration for thinking.type: "disabled".
2. Replace it with enabled.
3. Set reasoning_effort to low if the previous behavior was optimized for speed, or use high/max for harder tasks.
4. Change the model identifier to glm-5.3 only after the thinking configuration is valid.
5. Re-test latency, quota consumption, tool calls, and long-running task recovery in a staging environment.

This is a breaking request-level change. Sending the old disabled-thinking setting to GLM-5.3 causes the request to fail.

Availability & Access

• GLM Coding Plan: Available to all plan users at launch in ZCode, Claude Code, OpenCode, and other supported clients.
• ZCode: Includes Goal mode for planning, coding, testing, and verification, plus remote monitoring through WeChat or Feishu.
• Quota accounting: The new points system meters input, cached input, and output tokens separately. Calls outside weekday peak hours receive a 50% points discount; peak hours are 14:00–18:00 UTC+8, Monday through Friday.
• Open weights: Announced for two weeks after launch, pending safety evaluation and hardening. Do not plan production self-hosting until the official checkpoint and license files are actually published.

Pricing & Plans

GLM-5.3 access is included in the GLM Coding Plan. The current public monthly list prices are:

Annual billing currently lists effective monthly rates of $12.60, $56, and $117.60. ZCode advertises a launch-period 1.5× quota boost through August 31, 2026; combined with cache savings, Z.ai markets up to 180% effective standard quota. Treat this as a dated promotion, not a permanent plan feature.

The launch post does not publish a separate general API token price for GLM-5.3. Teams needing pay-as-you-go access should verify the current Z.ai API pricing page rather than applying older GLM model rates.

Best For

• Engineering teams delegating multi-hour implementation, debugging, and optimization tasks
• Coding-agent platforms that need better completion rates without simply increasing output tokens
• Security teams performing authorized vulnerability discovery and coordinated disclosure
• Developers already using ZCode, Claude Code, or OpenCode through the GLM Coding Plan
• Research groups evaluating post-training and long-horizon reinforcement-learning systems

FAQ