GLM-5.3

At a glance

Specification	Value
Provider	Z.ai
Family	GLM
Release date	2026-08-18 Vendor source
Context window	1,048,576 tokens Vendor source
Input modalities	text
Output modalities	text
Reasoning	Yes
Open weights	Yes
Borg lane ids	glm-5.3, z-ai/glm-5.3
Versions in Borg	GLM-4.5 · GLM-4.6 · GLM-4.7 · GLM-5 · GLM-5.1 · GLM-5.2 · GLM-5.3
Median time to first token (gateway proxy) on Borg	7293 ms
Median tokens per second (end to end) on Borg	25.2 output tokens per second

Not published by the vendor: thinking modes, knowledge cutoff, fusion roles.

What GLM-5.3 is good for

GLM-5.3 builds on the same base model as GLM-5.2, with improvements from post-training rather than a new base architecture. The important differences are stronger complex programming, terminal-agent work and security-oriented code analysis; unlike earlier siblings, reasoning is always on, with adjustable effort rather than a non-thinking switch. It retains a project-scale text context and large output allowance, making it useful for coding agents that must inspect a repository, operate tools and carry changes through to completion. Z.ai also reports gains on vulnerability-discovery tests, though an agent's findings still require independent verification. Choose it when quality on demanding engineering steps matters more than choosing a lightweight model. Borg includes GLM-5.3 in its subscription; Fusion can route to it, or you can select the lane yourself.

• complex coding-agent tasks
• terminal-based repository work
• security code review with verification
• deep reasoning on long engineering tasks

How Borg uses it

GLM-5.3 is available as a subscription catalog lane in Borg. The catalog does not publish which Fusion profiles route to this model, so no routing role is claimed here. You can pick it directly by setting the OpenAI-compatible gateway model field to a listed lane id. Check the current catalog for availability before building a workflow around a lane.

model: "glm-5.3"

Measured on Borg

Time to first token (gateway proxy)

Output tokens per second (whole turn)

• Time to first token (gateway proxy): time until the first streamed response event reaches the gateway; our closest available measure of time to first token
• Tokens per second (end to end): tokens per second, end to end: provider-reported output tokens divided by the whole turn's elapsed time, including queueing and reasoning; not a streaming per-token rate

Completion speed covers the whole turn including queueing and reasoning time; it is not a streaming tokens-per-second figure.

Measured on Borg: SearchAtlas AI Gateway production request logs. 30-day window ending 2026-10-08; 659807 requests. Minimum 100 samples per published request metric.

Benchmarks

Benchmark	Vendor-reported score	Source
Terminal-Bench 3.0	28.3	Vendor source
DeepSWE v1.1	66.9	Vendor source
Agents' Last Exam	28.5	Vendor source
CyberGym	84.5%	Vendor source
ExploitBench	54.4%	Vendor source
Z.ai Code Bench, Max effort completion	34.5%	Vendor source

Borg coding bench

Hard-graded with tests plus lint, no LLM judge; 3 tasks, 15 runs, 2026-09-20.

Source: Borg coding bench, hard-graded by tests and ruff; no LLM judge

Pricing

Included in Borg subscription. Vendor API list price: $1.4 input / $4.4 output per 1M tokens. Vendor price source.

Compare subscription and API pricing · Cost Savings Calculator

FAQ

What is GLM-5.3?

GLM-5.3 is listed in Borg's subscription catalog as glm-5.3. GLM-5.3 builds on the same base model as GLM-5.2, with improvements from post-training rather than a new base architecture.

Is GLM-5.3 in Borg?

Yes. GLM-5.3 is listed in Borg's subscription catalog under the lane glm-5.3. You can choose the lane directly, or explore the available Fusion profiles. Availability and model details can change; check the current catalog before deployment.

What context window does GLM-5.3 have?

The published context window is 1,048,576 tokens. Consult the cited sources for details and current limits. Vendor source

Is GLM-5.3 good for coding?

Vendor-published evaluations appear in the benchmarks section; their settings may differ from your tasks. Its listed uses include complex coding-agent tasks, terminal-based repository work, security code review with verification, deep reasoning on long engineering tasks. Try the lane for your own tasks, and compare the output with your requirements before choosing a default.

How fast is GLM-5.3 on Borg?

Measured on Borg: SearchAtlas AI Gateway production request logs. 30-day window ending 2026-10-08; 659807 requests. Minimum 100 samples per published request metric.

How does GLM-5.3 do on Borg's coding bench?

GLM-5.3 passed 53% overall and 53% of hidden tests across 15 runs.

How much does GLM-5.3 cost on Borg vs the API?

This model is included in the Borg subscription. The vendor API list prices appear below with a source link. See the pricing explainer and cost savings calculator to compare billing approaches.

Sources

• https://docs.z.ai/guides/llm/glm-5.3.md
• https://docs.z.ai/release-notes/new-released.md
• https://docs.z.ai/guides/overview/pricing.md
• https://huggingface.co/zai-org/GLM-5.3/raw/main/README.md

Related

Other versions of GLM: GLM-4.5 · GLM-4.6 · GLM-4.7 · GLM-5 · GLM-5.1 · GLM-5.2

Other Z.ai models: GLM-4.5 · GLM-4.5-Air · GLM-4.6 · GLM-4.7 · GLM-5 · GLM-5-Turbo · GLM-5.1 · GLM-5.2

Models sharing a Fusion role: Not published

All makers and models