Databricks Swaps Claude Opus 4.8 for Zhipu AI's GLM 5.2 and Cuts Costs 34%

On July 11, 2026, Databricks announced a strategic decision: it is making the open Chinese model GLM 5.2 from Zhipu AI its primary tool for automating coding. In tests on its own multimillion-line codebase, GLM 5.2 produced results identical to Anthropic's Claude Opus 4.8, but 34% cheaper: $1.28 per task versus $1.94.

How Databricks tested the models

Databricks engineers ran coding agents on the company's real production codebase, millions of lines of source code that reflect the full complexity and specifics of real-world development. This matters because public benchmarks often measure models on synthetic tasks, which don't match how a model behaves on your own code.

Databricks measured not only the quality and speed of the generated code but also the practical cost at scale: how much money actually goes into coding automation per day, week and month in production. That is where the open GLM 5.2 showed its strength.

Results: GLM 5.2 reaches the same level as Opus 4.8

The results are telling:

• GLM 5.2 from Zhipu AI: $1.28 per coding task
• Claude Opus 4.8 from Anthropic: $1.94 for the same task
• Savings from switching to GLM 5.2: $0.66 per task, a 34% cost reduction
• Quality of the generated code: identical
• Execution speed and accuracy: comparable
• Tests were run on Databricks production data, not on public benchmarks

Databricks tested other candidates too, but GLM 5.2 came out on top for price-to-quality ratio.

Why this challenges conventional wisdom

The industry's traditional view is that closed Western models (Claude Opus, OpenAI's GPT, Google's Gemini) automatically outperform open alternatives from Asia. Databricks punctures that myth in practice: on real production tasks, the open GLM 5.2 from China's Zhipu AI codes as well as the proprietary Opus 4.8, with noticeable savings.

"Companies often rely only on publicly available benchmarks and don't test models on their own data.

Results can differ significantly," is, in essence, Databricks' conclusion.

That raises the central question: if open models compete with closed ones in practice, why pay 34% more?

What this means for the market

The AI model market is moving away from the dominance of one or two providers toward real diversity. Open models from Zhipu AI (GLM), Meta* (Llama) and Mistral are becoming competitive not only on price but also on quality in practical production tasks. For companies this means more choice: there is no need to blindly pick the model with the best public ranking. Instead, test on your own data and calculate the real cost of ownership.

Frequently asked questions

When will GLM 5.2 be publicly available?

GLM 5.2 is already available as an open-source model from Zhipu AI. Any company can download it, test it and deploy it on its own servers or in the cloud without licensing fees.

How much would a company actually save by switching to GLM 5.2?

Savings depend on usage volume. If a company spends $100 a day on Claude Opus 4.8 for coding, switching to GLM 5.2 would save roughly $34 a day, or $12,400 a year. But this requires testing on your own codebase.

*Meta is designated an extremist organization and banned in Russia.