How much does GLM-5.3 cost?

API pricing follows GLM-5.2's structure: $1.40 per million input tokens, $4.40 per million output, with cached input at $0.26. The GLM Coding Plan subscription starts at $18/month, and calls outside weekday peak hours (14:00-18:00 UTC+8) consume 50% of standard points.

Are GLM-5.3 open weights available?

Not immediately. Unlike previous GLM releases, the open weights are staged rather than published on launch day (August 14, 2026). Z.ai says weights will follow; until then GLM-5.3 is API-and-subscription-only, which is a notable break from its open-weight-first reputation.

Quick Answer

What Is GLM-5.3? Z.ai's New Coding Model (Aug 2026)

Q: What is GLM-5.3?

GLM-5.3 is Z.ai's coding-focused frontier model, released August 14, 2026. It runs on the same 743B-parameter base as GLM-5.2 — every reported gain comes from scaled post-training (more task environments, longer training). The biggest jump is on long-horizon coding: Terminal-Bench 3.0 moves from 4.6 to 28.3.

Published: August 15, 2026

The Short Answer

GLM-5.3 is Z.ai’s coding-focused model released August 14, 2026. It is not a new architecture and not a new pretrain — it runs on the same ~743B-parameter base as GLM-5.2, and every reported gain comes from scaled post-training: more task environments, more environment types, longer training runs. The headline result is long-horizon coding: Terminal-Bench 3.0 jumps from 4.6 to 28.3.

Key Facts

	GLM-5.3
Released	August 14, 2026
Base model	Same ~743B base as GLM-5.2 (post-training only)
API price	$1.40 in / $4.40 out per MTok
Cached input	$0.26 per MTok
Coding Plan	From $18/month
Off-peak discount	50% of standard points outside 14:00-18:00 UTC+8 weekdays
Open weights	Staged — not available at launch
Thinking	Mandatory; three effort levels (low, high, max)

What Actually Improved

The gains concentrate on agentic, long-horizon work rather than one-shot Q&A:

Terminal-Bench 3.0: 4.6 → 28.3 — the largest single jump, on the hardest terminal-agent benchmark generation.
Cyber and automation: Z.ai reports GLM-5.3 leading CyberGym and AutomationBench among open-line models — enough of a step that its emergent cyber capability drew security commentary on launch week.
Complex coding: improvements on DeepSWE-style multi-file tasks, per Z.ai’s launch materials.

The lesson labs keep re-learning in 2026: post-training scale is a product lever on its own. Z.ai shipped a meaningfully better coding agent without touching the pretrain.

What Changed Operationally

Two changes matter for anyone deploying it:

Thinking is now mandatory. You pick an effort level (low/high/max) but cannot disable reasoning — budget output tokens accordingly.
Weights are staged, not immediate. GLM built its reputation on day-one open weights; GLM-5.3 launched API-first with weights “to follow.” If your stack depends on self-hosting, GLM-5.2 remains the current self-hostable GLM (see best open-weight coding model 2026).

How It Stacks Up

At $1.40/$4.40 per MTok, GLM-5.3 undercuts Western frontier coding rates (Claude Opus 4.8/Opus 5 at $5/$25; GPT-5.6 Sol at $5/$30) while claiming competitive agentic-coding scores. Against open-weight rivals it competes with Kimi K3 and DeepSeek V4 — with the caveat that, for now, only its rivals give you the weights.

Who Should Care

Coding-agent builders wanting near-frontier long-horizon performance at ~3-6× less than Western API rates.
GLM Coding Plan users — the $18/month plan now fronts a materially better model.
Self-hosters — wait for the staged weights before counting GLM-5.3 as an open model.

Last verified: August 15, 2026.