Databricks has made its choice: the Chinese open-source model GLM 5.2 will become the company's daily coding engine. The reason is pragmatic—in internal benchmarks on its own multi-million-line codebase, GLM 5.2 proved statistically equivalent to Anthropic's Opus 4.8, but at significantly lower cost. While Opus costs $1.94 per task, GLM 5.2 comes in at $1.28. It's a clear signal: no single provider dominates anymore.
The essentials
- GLM 5.2 and Opus 4.8 both achieve 82–90% pass rates in the top performance tier, but GLM 5.2 costs significantly less per task
- Databricks plans to deploy the model immediately as a "daily driver" for its developers
- Other US firms are following suit: Coinbase halved AI spending using Chinese models, Lindy achieved savings with Deepseek v4
- On OpenRouter, Chinese models now account for over 30% of weekly traffic since February 2026—up from 11% the year before
Internal benchmarks beat public rankings
Databricks emphasizes a crucial point: the company built its own benchmark using real, production tasks from actual work. Public datasets are often unrepresentative of a company's own codebase, and models can "cheat" through training data. The result: three clear performance clusters emerged.
| Cluster | Pass Rate | Examples |
|---|---|---|
| Top | 82–90% | Opus 4.8, GLM 5.2, GPT 5.5 |
| Mid | 71–82% | Sonnet 4.6, Sonnet 5, GPT 5.4 |
| Base | 51–60% | GPT 5.4-mini, Haiku 4.5 |
The insight from Databricks co-founder Matei Zaharia and team: > "The evidence shows it's time to start deploying these as daily drivers for coding." Internal pilots with developers confirmed the measurements.
Cost optimization through intelligent routing
Databricks also analyzed the complexity of its coding tasks: 61% are medium complexity, 19% low, only 12% high. Until now, all ran through the most expensive models. Going forward, the company plans to route simpler tasks to cheaper models—a classic tiering model that cuts costs without sacrificing quality.
The so-called Pareto frontier (best quality-to-cost ratio) is achieved, according to Databricks, only through a mix of OpenAI, Anthropic, and open-source models. No single provider dominates anymore.
The China trend is real
Databricks is not alone. Coinbase switched to Chinese models like GLM 5.2 and Kimi 2.7 and halved AI spending while token usage climbed. Lindy replaced Claude entirely with Deepseek v4. Snowflake tested GLM 5.2 against Opus 4.7—virtually tied, but a fraction of the cost.
On the AI routing platform OpenRouter, the trend is stark: Chinese models now make up over 30% of weekly traffic, with 60–90% lower costs than Western alternatives.
What this means for European enterprises
For European tech and data companies, this is a wake-up call. The assumption that only Anthropic, OpenAI, and Google build competitive models is being empirically disproven—at least for coding tasks. Companies should, as Databricks demonstrates, run their own benchmarks on real codebases rather than relying on public rankings. That saves costs and reveals which models actually fit your infrastructure. At the same time, questions of data security and vendor lock-in arise—topics European firms must take seriously given regulatory requirements.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
