Databricks Defaults to Chinese Model GLM 5.2, Matches Opus at $1.28/Task

#ai #opensource #programming #machinelearning

Databricks defaulted to GLM 5.2 after it matched Opus 4.8 at $1.28/task vs $1.94. The move signals enterprises building custom benchmarks and multi-vendor AI stacks.

Databricks benchmarked GLM 5.2 against Anthropic's Opus 4.8 on its own codebase and found them statistically tied. The Chinese open-source model costs $1.28 per task versus $1.94 for Opus.

Key facts

GLM 5.2 costs $1.28 per coding task vs Opus 4.8 at $1.94.
Top tier pass rate: 82–90% across Opus 4.8, GLM 5.2, GPT 5.5.
61% of Databricks coding tasks are medium complexity.
Coinbase cut AI spending in half using GLM-5.2 and Kimi 2.7.
Chinese models hit 30% of OpenRouter traffic by Feb 2026.

Databricks ran internal coding agent benchmarks on its multi-million-line codebase and found that the Chinese open-source model GLM 5.2 matched Anthropic's Opus 4.8 in pass rate. The cost difference — $1.28 per task versus $1.94 — has prompted the company to adopt GLM 5.2 as its default daily coding engine. According to The Decoder

The benchmark grouped models into three performance tiers. The top cluster, with an 82 to 90 percent pass rate, includes Opus 4.8, GLM 5.2, and certain configurations of GPT 5.5. A middle tier at 71 to 82 percent holds Sonnet 4.6, Sonnet 5, and GPT 5.4. The bottom tier at 51 to 60 percent includes GPT 5.4-mini and Haiku 4.5. Notably, no single provider dominates the top tier.

Key Takeaways

Databricks defaulted to GLM 5.2 after it matched Opus 4.8 at $1.28/task vs $1.94.
The move signals enterprises building custom benchmarks and multi-vendor AI stacks.

The Cost-Intelligence Frontier

Databricks co-founder Matei Zaharia co-authored the blog post, stating: "The evidence shows it's time to start deploying these as daily drivers for coding." Developer feedback from internal pilots backed the results. The company plans to route tasks based on complexity — 61 percent of coding tasks are medium complexity, 19 percent low, and only 12 percent high — sending cheaper models to simpler work.

Databricks is not alone in this shift. Coinbase moved to Chinese models including GLM-5.2 and Kimi 2.7, cutting AI spending in half while token usage kept climbing. Lindy ditched Claude entirely for Deepseek v4 and saved millions. Snowflake tested GLM-5.2 against Opus 4.7 and found them nearly tied at a fraction of the cost. On OpenRouter, Chinese models have topped 30 percent of weekly traffic since February 2026, up from 11 percent last year, at 60 to 90 percent lower cost than Western alternatives.

Why Public Benchmarks Fail

Databricks developed its own benchmark using real-world tasks because public datasets are often not representative of their codebase and models can "cheat" by leveraging prior knowledge from training data. The company argues that token price and actual task cost are not the same — token efficiency matters just as much. The Pareto frontier for quality-to-cost is shaped by models from three providers: OpenAI, Anthropic, and open source. Only a mix delivers frontier-level performance, Databricks says.

The broader takeaway: no single provider dominates, and companies should build their own benchmarks instead of relying on public ones. This aligns with a trend where enterprises increasingly treat AI procurement as a multi-vendor optimization problem rather than a single-platform bet.

What to watch

Watch for Q3 2026 enterprise adoption numbers for GLM-5.2 and whether other major tech companies follow Databricks in building custom internal benchmarks. Also monitor OpenRouter traffic share for Chinese models, which could cross 40% by year-end.