Z.ai Disrupts AI Landscape with GLM-5 An Open-Weights Model Rivaling GPT-5.2 and Gemini 3 Pro

China Z.ai Unleashes GLM-5: The Open-Source Challenger to GPT-5.2.

Chinese AI powerhouse Z.ai has officially released GLM-5, a groundbreaking open-weights model designed to compete directly with closed-source giants like GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. This new release marks a significant leap over its predecessor, GLM-4.7, particularly in autonomous reasoning and agentic tasks.

Dominating Agentic AI Benchmarks

GLM-5 has shown remarkable performance in benchmarks focused on Agentic AI. Notably:

• BrowserComp: GLM-5 outperformed all other current models, showcasing superior ability in navigating and executing tasks within web browsers.

• MCP-Atlas: The model delivered scores nearly identical to GPT-5.2, proving its competence in complex tool-use and multi-step reasoning.

Architecture and Innovation

The model’s architecture has seen a massive scale-up. GLM-5 features a total of 744B parameters with 40B active parameters, compared to GLM 4.5's 355B total and 32B active parameters. This efficiency is achieved through an advanced Asynchronous Reinforcement Learning (RL) training process, which optimizes the model's decision-making capabilities without traditional bottlenecks.

Pricing and Availability

Reflecting its enhanced capabilities, GLM-5 comes with a higher price tag roughly 66% more expensive than GLM-4.7:

• API Pricing: $1.00 per million input tokens and $3.20 per million output tokens.

• Subscription Access: Currently, GLM-5 is exclusive to GLM Coding Plan "Max" subscribers ($80/month). Z.ai plans to roll out access to "Pro" tier users in the near future.

The fact that GLM-5 has a total of 744 billion parameters but only uses 40 billion actively reflects its highly advanced Sparse MoE technology. This enables the model to be super-intelligent while still running at acceptable speeds and with lower energy efficiency than typical dense models.

Asynchronous RL training allows the model to learn from errors in parallel, giving GLM-5 a more human-like reasoning ability and reducing hallucinations in complex coding and computational tasks.

Releasing it as open-weighted while having capabilities equivalent to GPT-5.2 is Z.ai's "rural encirclement of urban areas" strategy to attract developers worldwide to use the GLM architecture instead of relying solely on the closed US system.

Even though the price has increased by 66%, analysts consider it worthwhile for organizations requiring autonomous agents because GLM-5's success rate is significantly higher than its predecessor, reducing retries and ultimately saving overall costs.

 

 

The $20K Experiment: How Anthropic's Researcher Used 16 AI Agents to Compile Linux.

 

Source: Z.ai 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

China Z.ai Unleashes GLM-5: The Open-Source Challenger to GPT-5.2.

Chinese AI powerhouse Z.ai has officially released GLM-5, a groundbreaking open-weights model designed to compete directly with closed-source giants like GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. This new release marks a significant leap over its predecessor, GLM-4.7, particularly in autonomous reasoning and agentic tasks.

Dominating Agentic AI Benchmarks

GLM-5 has shown remarkable performance in benchmarks focused on Agentic AI. Notably:

• BrowserComp: GLM-5 outperformed all other current models, showcasing superior ability in navigating and executing tasks within web browsers.

• MCP-Atlas: The model delivered scores nearly identical to GPT-5.2, proving its competence in complex tool-use and multi-step reasoning.

Architecture and Innovation

The model’s architecture has seen a massive scale-up. GLM-5 features a total of 744B parameters with 40B active parameters, compared to GLM 4.5's 355B total and 32B active parameters. This efficiency is achieved through an advanced Asynchronous Reinforcement Learning (RL) training process, which optimizes the model's decision-making capabilities without traditional bottlenecks.

Pricing and Availability

Reflecting its enhanced capabilities, GLM-5 comes with a higher price tag roughly 66% more expensive than GLM-4.7:

• API Pricing: $1.00 per million input tokens and $3.20 per million output tokens.

• Subscription Access: Currently, GLM-5 is exclusive to GLM Coding Plan "Max" subscribers ($80/month). Z.ai plans to roll out access to "Pro" tier users in the near future.

The fact that GLM-5 has a total of 744 billion parameters but only uses 40 billion actively reflects its highly advanced Sparse MoE technology. This enables the model to be super-intelligent while still running at acceptable speeds and with lower energy efficiency than typical dense models.

Asynchronous RL training allows the model to learn from errors in parallel, giving GLM-5 a more human-like reasoning ability and reducing hallucinations in complex coding and computational tasks.

Releasing it as open-weighted while having capabilities equivalent to GPT-5.2 is Z.ai's "rural encirclement of urban areas" strategy to attract developers worldwide to use the GLM architecture instead of relying solely on the closed US system.

Even though the price has increased by 66%, analysts consider it worthwhile for organizations requiring autonomous agents because GLM-5's success rate is significantly higher than its predecessor, reducing retries and ultimately saving overall costs.

 

 

The $20K Experiment: How Anthropic's Researcher Used 16 AI Agents to Compile Linux.

 

Source: Z.ai