GLM-4.5

// Overview

GLM-4.5 is the flagship large language model released by Z.ai (formerly Zhipu AI) on 28 July 2025. It is a Mixture-of-Experts model with 355 billion total parameters and 32 billion active per forward pass, shipped under a permissive MIT license with open weights — base, hybrid (thinking/non-thinking) and FP8 variants are all available for commercial use and fine-tuning. A smaller GLM-4.5-Air sibling (106B total / 12B active) targets cheaper, lighter deployment.

GLM-4.5 was designed as an agent-native model: a single set of weights unifies reasoning, coding and tool use rather than splitting them across separate models. Z.ai reports a tool-calling success rate of 90.6% and positions the model for software engineering, web browsing and front-end development. A hybrid reasoning mode lets callers switch between a deliberate thinking mode for hard tasks and a fast non-thinking mode for immediate answers.

On Z.ai's own aggregate of 12 representative benchmarks, GLM-4.5 scored an average of 63.2, which the company reported as third overall and the top score among open-source models at launch. It supports a 128K-token context window and up to 96K output tokens, and is served through the Z.ai API as well as third-party hosts such as OpenRouter.

// Benchmarks

1. MMLU-Pro84.6%
2. AIME 202491%
3. SWE-bench Verified64.2%
4. TAU-bench70.1%
5. BrowseComp26.4%
6. Tool-calling success rate90.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Standard Z.ai API list price for GLM-4.5 (the GLM-4.5-Air variant is cheaper at $0.20 input / $1.10 output). Z.ai ran a lower promotional launch rate; third-party hosts may differ.

Pricing source ↗

// Strengths

• Open weights under a permissive MIT license — free for commercial use, fine-tuning and on-prem deployment
• Agent-native design with a high reported tool-calling success rate (90.6%)
• Strong agentic and coding scores (SWE-bench Verified 64.2, TAU-bench 70.1) for an open model
• Hybrid thinking / non-thinking modes in one model, so you trade off depth vs. latency per request
• Aggressive pricing relative to closed frontier models, plus a lighter GLM-4.5-Air variant

// Best for

• Autonomous and tool-using agents (function calling, web browsing, multi-step workflows)
• Software engineering and front-end development assistance
• Math and competition-style reasoning tasks
• Self-hosted or fine-tuned deployments where open weights and an MIT license are required
• Cost-sensitive production workloads that still need frontier-class reasoning

// How to access

// 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.