Overview
GLM-5.3 is the GLM flagship Z.ai launched on August 14, 2026. Unlike the jump from GLM-5.1 to GLM-5.2, it is not a new pretraining run: Z.ai's own model documentation says it 'uses the same base model as GLM-5.2 — all improvements come from post-training'. It is positioned as a Flagship Foundation Model, takes text in and returns text, and carries a 1M-token context window with up to 128K output tokens.
The headline gains are in software engineering. Z.ai reports GLM-5.3 improving 50% over GLM-5.2 on its internal Z.ai Code Bench and reaching open-source state of the art on Terminal-Bench 3.0 and Agents' Last Exam (CLI). The figures it publishes for the two generations are Terminal-Bench 3.0 rising from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and Agents' Last Exam from 23.8 to 28.5. On GDPval-AA v2, which spans 44 occupations, Z.ai reports a score of 1769 points — an index figure rather than a percentage, so it is not charted above.
Z.ai attributes the gains to post-training on complete engineering workflows rather than isolated programming problems: identifying a problem, analysing the solution, implementing, verifying and delivering, with some training tasks equivalent to several days of a senior engineer's work and requiring real compute clusters, storage systems and internal code. The company says the model can carry projects across tens of thousands of lines of code and hundreds of files with minimal human intervention.
The second, less expected result is cybersecurity. Z.ai reports 84.5% on CyberGym for vulnerability discovery — ahead of the 83.8% it lists for Mythos 5 and the 83.6% for GPT-5.6 Sol — and ExploitBench rising from 24.4% to 54.4%. Z.ai qualifies this: the advantage 'mainly lies in the front end of the vulnerability exploitation chain, and there is still room for improvement in more in-depth vulnerability exploitation and complete offensive and defensive tasks'.
Availability at launch was through the GLM Coding Plan subscription; the model API followed. Z.ai's documentation lists GLM-5.3 as fully available over three protocols — an OpenAI Chat Completions endpoint, an OpenAI Responses endpoint and an Anthropic Messages endpoint — and its pricing page lists the model at $1.40 per million input tokens, $0.26 cached input and $4.40 output, matching GLM-5.2's rates. Z.ai notes that accounts whose only entitlement is an active or expired GLM Coding Plan can reach the API through the OpenAI Chat Completions-compatible protocol alone.
The weights are a separate track. Z.ai's launch post says they will be published about two weeks after launch, once safety evaluation and hardening are complete. No GLM-5.3 weights had appeared in the zai-org Hugging Face organisation as of August 19, 2026, so the model is served through Z.ai's API rather than self-hosted, and carries no published licence.
| Released | 2026-08-14 |
|---|---|
| License | Not published at launch — Z.ai said the GLM-5.3 weights and their licence would follow about two weeks after the August 14, 2026 launch, once safety evaluation and hardening were complete |
| Weights | API only |
| Parameters | Same base model as GLM-5.2 (744B total · 40B active) |
| Context | 1M |
| Max output | 128K |
| Architecture | Mixture-of-Experts. Z.ai states GLM-5.3 'uses the same base model as GLM-5.2 — all improvements come from post-training', so the pretrained stack is unchanged and every reported gain comes from the post-training stage. |
| Modalities | Text |
| Status | Available |
Benchmarks
Z.ai's "Performance across comparison models" table, transcribed from the GLM-5.3 launch post; the column names are Z.ai's own. Blank cells are figures Z.ai did not publish. ExploitGym counts tasks completed under time-normalised 2-hour / 6-hour budgets and GDPval-AA v2 is an index score measured by Artificial Analysis — neither is a percentage.
| Benchmark | GLM-5.3 | GLM-5.2 | Kimi K3 | DeepSeek-V4 Pro-0813 | Qwen3.8-Max | Opus 4.8 | Fable 5 (w/ fallback) | GPT-5.6 Sol |
|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 88.2% | 81% | 88.3% | 87.9% | 86.6% | 85% | 88% | 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% | 56.6% | 58% | 69.7% | 72.7% |
| NL2Repo | 58% | 48.9% | 58% | 61.1% | 55.9% | 69.7% | — | — |
| ProgramBench (Almost Solved) | 19% | 9.5% | 17.5% | — | 10.5% | 15.5% | 33% | 23% |
| FrontierSWE | 78.1% | 67.5% | — | — | — | 66.5% | 88.2% | — |
| SWE-Marathon (v1.1) | 42.5% | 19.4% | 48.1% | — | — | 48.8% | 33.1% | 42.5% |
| PostTrainBench | 39.8% | 31.7% | 32% | — | — | 32.9% | 41.8% | 36.2% |
| CyberGym | 84.5% | 77.2% | 80% | 83.3% | 78.5% | 78.1% | 83.8% | 83.6% |
| ExploitGym (2h / 6h tasks) | 105 / 130 | 29 / 39 | 36 / 70 | — | 14 / 26 | 80 / 120 | 181 / 247 | 216 / 293 |
| ExploitBench | 54.4% | 24.4% | 32.2% | — | 28.8% | 40% | 78% | 76.5% |
| Toolathlon Verified | 73% | 59.9% | 76.5% | 74.1% | 72.5% | 76.2% | 74.7% | 74.9% |
| AutomationBench (v1.0.6) | 48.2% | 26.2% | 46.7% | 43.2% | 39.8% | 41% | 46.2% | 45.8% |
| Agents' Last Exam (ALE-CLI) | 28.5% | 23.8% | 27.6% | 25.7% | 27% | 25.7% | 23.8% | 28.6% |
| HLE w/ Tools | 62.5% | 54.7% | 59.8% | 60% | 56.2% | 57.9% | 63.9% | 64.5% |
| GDPval-AA v2 | 1769 pts | 1508 pts | 1682 pts | 1590 pts | 1739 pts | 1588 pts | 1743 pts | 1730 pts |
This model's scores
- Terminal-Bench 3.028.3%
- DeepSWE v1.166.9%
- Agents' Last Exam28.5%
- CyberGym (vulnerability discovery)84.5%
- ExploitBench54.4%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $1.40 / 1M tokens |
|---|---|
| Cached input | $0.26 / 1M tokens |
| Output | $4.40 / 1M tokens |
Z.ai's published list price, identical to GLM-5.2. Cached-input storage is listed as free for a limited time. The GLM Coding Plan bills separately on a points quota, with off-peak calls costing 50% of standard points.
Strengths
- Terminal-Bench 3.0 of 28.3 against 4.6 for GLM-5.2 — the largest single jump Z.ai published for this release
- DeepSWE v1.1 of 66.9 (GLM-5.2: 46.2) for long-horizon agentic software engineering
- CyberGym 84.5% for vulnerability discovery, ahead of the Mythos 5 and GPT-5.6 Sol figures Z.ai lists
- ExploitBench of 54.4%, more than double GLM-5.2's 24.4%
- 1M-token context with 128K max output, plus function calling, structured output, context caching and MCP tool use
- Delivered without a new pretraining run — the same base model as GLM-5.2, improved by post-training alone
Best for
- Long-horizon coding agents that must stay productive across hundreds of files and interdependent systems
- White-box code review, vulnerability discovery and verification workflows
- Terminal and CLI agent work, where Z.ai reports open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam
- Professional task execution beyond programming, per Z.ai's GDPval-AA v2 result across 44 occupations
How to access
| Provider | Model ID |
|---|---|
| Z.ai ↗ | glm-5.3 |
GLM (flagship) — every version
The full lineage of the GLM (flagship) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
FAQ
When was GLM-5.3 released?
Z.ai launched GLM-5.3 on August 14, 2026, first to GLM Coding Plan subscribers. The model API followed and is documented as fully available over OpenAI Chat Completions, OpenAI Responses and Anthropic Messages protocols.
How is GLM-5.3 different from GLM-5.2?
It runs on the same base model. Z.ai states that all of GLM-5.3's improvements come from post-training, not a new pretraining run — the post-training set was widened from isolated programming problems to complete engineering workflows spanning problem identification, implementation, verification and delivery.
Are GLM-5.3 weights available?
Not at launch. Z.ai said the weights would be published about two weeks after the August 14, 2026 launch, once safety evaluation and hardening were complete. No GLM-5.3 weights had appeared in the zai-org Hugging Face organisation as of August 19, 2026, so the model is served through Z.ai's API rather than self-hosted.
How much does the GLM-5.3 API cost?
Z.ai's pricing page lists GLM-5.3 at $1.40 per million input tokens, $0.26 per million cached input tokens and $4.40 per million output tokens — the same rates as GLM-5.2 — with cached-input storage free for a limited time. The GLM Coding Plan is billed separately on a points-based quota in which off-peak calls, including all day at weekends, consume half the standard points.
What is GLM-5.3's context window?
1M tokens, with up to 128K output tokens. Z.ai's documentation also lists thinking modes, streaming output, function calling, context caching, structured output and MCP tool integration.
How good is GLM-5.3 at security work?
Z.ai reports 84.5% on CyberGym for vulnerability discovery — ahead of the 83.8% it lists for Mythos 5 and 83.6% for GPT-5.6 Sol — and ExploitBench rising from 24.4% to 54.4%. Z.ai notes the advantage is concentrated at the front end of the exploitation chain, with deeper exploitation and full offensive/defensive tasks still weaker.