Databricks makes Chinese open-source model GLM 5.2 its default coding engine after it matched Opus at lower cost

2026-07-09 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, short

Summary

Databricks announced on July 9, 2026, its decision to make the Chinese open-source model GLM 5.2 its default coding engine after internal benchmarks revealed it statistically matched Anthropic's Opus 4.8 in performance but at a significantly lower cost. The company's analysis, conducted on its multi-million-line codebase, found GLM 5.2 cost \$1.28 per task compared to Opus's \$1.94. This move aligns with similar shifts by Coinbase, Lindy, and Snowflake, who have also adopted Chinese AI models for cost savings. Databricks' custom benchmark, designed to avoid public dataset biases and model "cheating," categorized models into three performance tiers, with GLM 5.2, Opus 4.8, and GPT 5.5 in the top group (82-90% pass rate). The company also highlighted that token efficiency, exemplified by its Pi harness reducing context and costs by up to 2.08x, is crucial for optimizing actual task expenses.

Key takeaway

For AI Engineering Directors evaluating coding assistant LLMs, you should prioritize custom internal benchmarks over public datasets to accurately assess model performance and cost-efficiency for your specific codebase. Consider integrating high-performing, lower-cost open-source models like GLM 5.2, especially for medium and low-complexity tasks, to significantly reduce operational expenses. Additionally, explore context-reduction techniques, such as Databricks' Pi harness, to optimize token usage and further improve your quality-to-cost ratio across your model portfolio.

Key insights

Open-source models like GLM 5.2 can match top proprietary LLMs in coding performance at significantly lower costs.

Principles

• Internal benchmarks reveal true model utility.
• Cost-performance Pareto frontier is diverse.
• Token efficiency impacts real-world cost.

Method

Develop custom benchmarks using recent, human-written, full-stack tasks with high-quality tests. Manually review tasks, rewrite tests for alternatives, and truncate Git history to prevent model "cheating."

In practice

• Route coding tasks to cheaper models by complexity.
• Implement context-reducing harnesses like Pi.
• Evaluate open-source alternatives for cost savings.

Topics

• GLM 5.2
• LLM Benchmarking
• Coding Assistants
• Cost Optimization
• Open-source AI
• Databricks

Best for: CTO, VP of Engineering/Data, Machine Learning Engineer, AI Engineer, Software Engineer, Director of AI/ML

Related on AIssential

• Open source llm (glm 4.7) matching closed models on coding benchmarks. Tested via api on real projects. — Open-source models are achieving competitive, specialized performance in domains like coding at significantly lower costs.
• Benchmarking Coding Agents on Databricks’ Multi-Million Line Codebase — Benchmarking coding agents on real-world codebases reveals optimal performance requires diverse models and efficient harnesses, not just low token costs.
• GLM 5.2: NEW Opensource KING IS BEATING GPT-5.5 & Opus 4.8! (Fully Tested) — GLM 5.2 is a new open-source model matching or exceeding proprietary LLMs in specific creative and coding tasks at a fraction of the cost.
• GLM 5.2 Z.ai Coding Model. From Beginner to Power User. Locally & API & Benchmark — GLM 5.2 offers a cost-effective, high-performance coding model comparable to leading LLMs like Fable 5.
• GLM 5.2 is SO GOOD (and almost free) — GLM 5.2 provides high-intelligence, open-weight model capabilities at a fraction of commercial API costs, enabling self-hosting and vendor independence.
• GLM 5.2: why I’m replacing Opus in Claude Code with this new model — GLM 5.2 offers high-intelligence, cost-effective, open-weight model capabilities comparable to leading proprietary LLMs for coding tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.