GLM 5.2 Wins Over Developers After Anthropic Fable Ban

Chinese open-source AI models are having their moment in the sun, countering US chip export restrictions with open-source accessibility and cost-effective performance. Now, a new reasoning model threatens to tilt the balance in China’s favour.

Last week, Chinese AI startup Z.ai (formerly Zhipu AI) released GLM 5.2, a new open-source large language model designed for running long coding tasks and agentic workflows.

The model has quickly gained attention for its performance in coding, reasoning, and agentic workflows. The reaction has been strong enough that several AI researchers and developers are calling it one of the most important open-model releases since DeepSeek R1.

The model also comes at a time when the US government has restricted access to Anthropic’s Claude Fable 5 and Claude Mythos 5, the AI giant’s strongest frontier models yet, excelling in autonomous software engineering, in-depth research, and high-level analytical tasks. This has prompted many non-US companies to look eastwards.

At the same time, Z.ai co-founder Jie Tang recently claimed that the company will build a Mythos-class model before the end of the year.

The enthusiasm surrounding GLM-5.2 is extending beyond the developer community. Bloomberg recently reported that Z.ai is weighing a multi-billion-dollar fundraising in Hong Kong after its shares rallied about 2,000% this year.

Z.ai is now widely regarded as one of China’s leading ‘AI Tigers’—a group of six domestic startups competing with Western AI leaders, including OpenAI, Anthropic, and Google. 

The GLM 5.2 moment is the new DeepSeek R1 moment.

The last remaining moat is gone, unless the US labs pull something unseen before from the sleeve.

— BURKOV (@burkov) June 21, 2026

A Competitive Model

GLM 5.2 builds on the GLM-5 series and uses a Mixture-of-Experts (MoE) architecture with 744 billion total parameters, of which roughly 40 billion are activated for each token during inference.

The model comes with a 1 million-token context window, support for long-running agent workflows, and new reasoning modes for complex coding tasks. 

According to Z.ai, the latest model introduces substantial improvements in reasoning and software engineering performance.

It performs competitively against closed models such as Claude Opus 4.8 and GPT-5.5 on coding benchmarks, including Terminal-Bench, SWE-Bench, and agentic evaluations, often narrowing the performance gap while operating at a fraction of the cost.

Beyond performance, GLM 5.2’s pricing has become a major talking point among developers. The model costs $1.40 per million input tokens and $4.40 per million output tokens, compared to $5 and $25, respectively, for Claude Opus 4.8, making it roughly 3.5x cheaper on input tokens and nearly 6x cheaper on output tokens.

GLM 5.2 is available both as an API service and as an open-weight model under the MIT license, allowing developers to either access it through hosted endpoints or run it on their own infrastructure. 

Developers can access it through Z.ai’s API platform, Hugging Face, OpenRouter, and Together AI. The model can also be integrated into coding environments such as Claude Code, Cline, and OpenClaw, making it relatively easy to test within existing developer workflows. 

The World Takes Notice

The excitement around GLM 5.2 has been driven less by marketing and more by word of mouth.

“GLM is the kind of model that revives serious interest in open source AI. It passes the blind test relative to the frontier models on the median production grade knowledge worker task,” wrote Perplexity AI CEO Aravind Srinivas in a post on X.  

He said the model is affordable to serve and, despite having fewer than 1 trillion parameters, has the potential to match frontier models on median-difficulty tasks while extending that performance to long-tail challenges.

Similarly, Yuchen Jin, an engineer at Databricks, said his early experience with GLM 5.2 suggested that the model frequently performed at a level comparable to Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5. “I compared it side by side with Opus 4.8, and sometimes I even preferred GLM-5.2’s results,” he wrote on X.

Mat Velloso, former Vice President at Meta and Google DeepMind, said that after spending an entire day using GLM 5.2, he barely felt the need to switch to another model, describing it as the first open model that meets his standards for daily use. “First open model that passes the bar as a daily driver. Things are not going to be the same. Damn, now I want to buy some serious hardware,” he wrote on X.

According to independent benchmarking platform Artificial Analysis, GLM 5.2 is currently the highest-ranked open-weight model on the GDPval-AA benchmark, which evaluates AI systems on real-world, economically valuable knowledge work through long-horizon, multi-turn tasks. The model scored 1,524 Elo, placing it third overall behind Claude Fable 5 and Claude Opus 4.8, while performing on par with GPT-5.5.

Guillermo Rauch, Founder and CEO of cloud platform Vercel, said he was ‘almost shocked’ by GLM 5.2’s coding performance, saying the model’s capabilities could change how developers view open-source AI.

Meanwhile, Chris Saum, Partner and Investor at Active Capital, said GLM 5.2 had become a compelling alternative to Anthropic’s models, noting that after switching from Claude, his daily AI spending fell from around $300 to just $3.82. He also said the model was able to identify and fix a bug that Claude had missed the previous day. 

Win for Open-Source 

For much of the generative AI boom, the most capable models came primarily from US companies. OpenAI, Anthropic, and Google set the pace while open-source alternatives struggled to keep up. That gap appears to be narrowing.

Following the success of DeepSeek, Kimi, Qwen, and now GLM 5.2, Chinese AI labs are increasingly producing models that can compete with frontier systems.

Tang said that the path to AGI (Artificial General Intelligence) must never be enclosed by high walls. “We have always believed that AGI should be the cornerstone for all of humanity to collaboratively explore the boundaries of intelligence and solve complex challenges, rather than a privilege monopolised by a few and subject to revocation at any moment,” he wrote on X.

“In the face of external blockades and restrictions, our attitude is one of radical openness. Frontier intelligence must remain open-source, accessible, and buildable, serving every dedicated developer.”