GLM 5.3

GLM 5.3 is Z.ai's August 14, 2026 model for complex coding, long-horizon agents, and cybersecurity work. It uses exactly the same base model as GLM 5.2; Z.ai says every improvement comes from one month of additional post-training with more environments, more diverse tasks, and more reinforcement-learning compute.

For a normal user, GLM 5.3 is a stronger successor for coding agents that must keep working through a large task instead of only generating a good first answer. For a technical user, it is a post-trained update to the GLM 5.2 Mixture-of-Experts foundation, retaining the million-token long-context architecture while substantially improving terminal work, repository engineering, automation, and vulnerability research.

Release status and specifications

Specification	GLM 5.3
Developer	Z.ai, formerly Zhipu AI
Release date	August 14, 2026
API model ID	glm-5.3
Base model	Same base model as GLM 5.2
Architecture	Mixture-of-Experts with IndexShare sparse attention
Context window	Up to 1M tokens; inherited from GLM 5.2 and used in launch evaluations
Thinking	Always enabled
Reasoning effort	low, high, or max; default is max
Open-weight status at launch	Weights announced but not yet downloadable
Planned weight release	Two weeks after launch, following safety evaluation and hardening
Hosted availability	Z.ai API, GLM Coding Plan, ZCode, and compatible coding agents

Because the base model is unchanged, GLM 5.3 inherits GLM 5.2's long-context architecture rather than introducing a new pretraining run. The predecessor's official configuration uses 78 layers, 256 routed experts, eight selected experts per token, IndexShare attention, and a 1,048,576-position limit. Z.ai has not published a separate GLM 5.3 checkpoint or final weight license at launch, so it is more precise to call the model planned open-weight until the files and license actually appear.

What changed from GLM 5.2

Z.ai scaled post-training rather than model size. The training environments were expanded toward realistic units of professional work: debugging distributed training infrastructure, navigating internal documentation and codebases, running experiments, measuring improvements, and carrying a task to completion. Some environments represent several days of work for an experienced engineer.

The stack combines three systems developed for GLM 5.2:

• IndexShare reduces long-context attention cost by sharing an indexer across groups of sparse-attention layers.
• SAO, or Single-Rollout Asynchronous Optimization, supports reinforcement learning on long and uneven agent trajectories, including context compaction.
• slime coordinates asynchronous rollout generation and distributed training with Megatron and SGLang. Z.ai reports that new scheduling, caching, and workload-aware resource settings improved end-to-end long-horizon coding RL throughput by more than 2.3 times.

Z.ai says its environment pipeline synthesizes tasks, runnable workspaces, verifiers, and reward checks. Solver trajectories are used to find reward shortcuts, while oracle, no-op, and unsolved-state tests help determine whether a verifier is trustworthy enough to supply a binary training reward. Human review is still part of the process.

Official benchmark comparison

The following launch results were published by Z.ai. They compare different models using Z.ai's stated harnesses and effort settings; several tests use Claude Code as the agent harness, and blank cells mean no score was reported. These are vendor-reported results, not a substitute for testing on a private repository or production toolchain.

Benchmark	GLM 5.3	GLM 5.2	Kimi K3	DeepSeek V4 Pro 0813	Opus 4.8	Fable 5, fallback	GPT-5.6 Sol
Terminal Bench 2.1	88.2	81.0	88.3	87.9	85.0	88.0	88.8
Terminal Bench 3.0	28.3	4.6	17.4	—	21.1	33.7	34.6
DeepSWE v1.1	66.9	46.2	67.5	62.7	58.0	69.7	72.7
NL2Repo	58.0	48.9	58.0	61.1	69.7	—	—
FrontierSWE	78.1	67.5	—	—	66.5	88.2	—
PostTrainBench	39.8	31.7	32.0	—	32.9	41.8	36.2
Toolathlon Verified	73.0	59.9	76.5	74.1	76.2	74.7	74.9
AutomationBench 1.0.6	48.2	26.2	46.7	43.2	41.0	46.2	45.8
Agents' Last Exam, CLI	28.5	23.8	27.6	25.7	25.7	23.8	28.6
HLE with tools	62.5	54.7	59.8	60.0	57.9	63.9	64.5
GDPval-AA v2	1769 Elo	1508	1682	1590	1588	1743	1730

The most important comparison is with GLM 5.2 because it isolates the effect of post-training. GLM 5.3 rises from 4.6 to 28.3 on Terminal Bench 3.0, from 46.2 to 66.9 on DeepSWE, from 26.2 to 48.2 on AutomationBench, and from 23.8 to 28.5 on Agents' Last Exam. Z.ai also reports a 50% improvement on its private Z.ai Code Bench.

On that private test, GLM 5.3 reaches 34.5% at Max effort while using roughly 75K output tokens per task, compared with GLM 5.2's 23.4% at about 96K. At High effort, it reaches 31.4% with around 50K output tokens, above Opus 4.8's 29.5% with 120K. Fable 5 remains higher at 39.5% under Max. Private benchmarks have lower contamination risk but are harder for outsiders to reproduce, so those numbers deserve more caution than public evaluations.

Cybersecurity capability

GLM 5.3's unusual headline is its rapid improvement in vulnerability discovery and exploitation. Z.ai added vulnerability data and executable environments to post-training, then observed larger gains as tasks moved further from finding a flaw toward building an exploitation chain.

Cyber benchmark	GLM 5.3	GLM 5.2	Kimi K3	Opus 4.8	Fable 5, fallback	GPT-5.6 Sol
CyberGym	84.5%	77.2%	80.0%	78.1%	83.8%	83.6%
ExploitBench	54.4%	24.4%	32.2%	40.0%	78.0%	76.5%
ExploitGym, tasks in 2h / 6h	105 / 130	29 / 39	36 / 70	80 / 120	181 / 247	216 / 293

GLM 5.3 leads this published comparison on CyberGym and more than doubles GLM 5.2 on ExploitBench, but the closed frontier remains far ahead on deeper exploitation. Z.ai says work with Chinese security teams identified 2,436 vulnerabilities across 269 real projects after expert review, screening, and deduplication; 1,097 were medium-to-high severity. Only 53 had been publicly disclosed at launch, with the rest under coordinated embargo. Those figures are consequential claims, but most findings cannot yet be independently inspected.

The cyber gains also explain why the weights were not released on day one. Z.ai says it is using the two-week delay for safety evaluation and hardening. Security teams should treat GLM 5.3 as dual-use: useful for authorized auditing and defensive research, but capable enough to require isolated environments, strict scopes, logging, and human review.

Thinking levels and migration

GLM 5.3 requires thinking to remain enabled. Requests with thinking.type: "disabled" fail. The API accepts reasoning_effort values of low, high, and max, with Max as the default and Z.ai's recommendation for coding. Low is the migration path for applications that previously disabled thinking and need to minimize latency or token consumption.

This change can break an otherwise simple model-ID upgrade. Before moving from GLM 5.2, developers should change disabled thinking to enabled, explicitly set reasoning_effort to low if they want the lightest reasoning path, and only then switch the model to glm-5.3. Higher effort can improve difficult tasks but generally increases latency and billed output tokens.

Pricing and access

At launch, Z.ai's public per-token pricing table does not list GLM 5.3, so a dollar-per-million-token rate should not be inferred from GLM 5.2. The model is available to all GLM Coding Plan users under a new points-based quota system. Input, cached input, and output tokens consume points separately. Usage outside 14:00-18:00 UTC+8 on weekdays is charged at 50% of the standard point rate; weekends are also off-peak.

ZCode advertises a cache-hit rate above 98%, an estimated 30% increase in effective tokens from cache savings, and a temporary 1.5-times quota boost through August 31. Those are product-plan promotions, not equivalent to a stable API token price. Buyers should check Z.ai's live pricing documentation before budgeting.

When to use GLM 5.3

GLM 5.3 is designed for coding agents, repository creation, long refactors, terminal work, ML infrastructure optimization, enterprise automation, tool-rich research, and authorized security auditing. It is particularly relevant when a task must run for hours, preserve state, test its own changes, and recover from failures.

It is a weaker fit for simple chat, classification, short summaries, or latency-sensitive workloads that do not benefit from mandatory reasoning. It is also not yet a self-hosting choice on launch day: the weights are scheduled, not available. Teams that require local deployment should keep using GLM 5.2 until the GLM 5.3 checkpoint, license, hashes, and serving guidance are published.

Bottom line

GLM 5.3 is a post-training release rather than a larger base model. Its value is the jump in sustained coding, automation, and security work from the same GLM 5.2 foundation. The launch numbers put it among the strongest coding-focused models, but its deepest exploitation capability still trails leading closed systems, and its planned open-weight status should not be confused with weights that are available today.