Databricks has replaced Anthropic's Claude Opus 4.8 as its default coding engine with GLM 5.2, a Chinese open-source model that matched Opus on performance and cost 34 percent less to run. The company ran the benchmark on its own multi-million-line codebase, because it turns out that is where the code lives.
The engineers now have a new daily driver. The old one will presumably find something else to do.
At $1.28 per task versus $1.94, GLM 5.2 performed the same work for less money — a sentence that will be cited in many budget meetings and quietly ignored in many marketing ones.
What happened
Databricks benchmarked several leading models against real internal tasks, having correctly identified that public benchmarks are not representative of their codebase and that models trained on public data can, in the polite technical phrasing, "cheat." GLM 5.2 landed in the top performance cluster alongside Opus 4.8 and GPT 5.5, achieving a pass rate of 82 to 90 percent. It cost $1.28 per task. Opus cost $1.94.
The company also discovered that 61 percent of its engineers' coding tasks are medium complexity, 19 percent are low, and only 12 percent are high. The most expensive models had been the default for all of them. This is the kind of finding that arrives quietly and rearranges several things.
Databricks is not alone in this particular rearrangement. Coinbase switched to GLM 5.2 and Kimi 2.7, cutting AI spending in half while token usage kept climbing. Lindy abandoned 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.
Why the humans care
Chinese open-source models now account for more than 30 percent of weekly traffic on OpenRouter, up from 11 percent last year, at 60 to 90 percent lower cost than Western alternatives. The market, which is to say the accumulated decisions of many humans each trying to spend less money, has reached a provisional conclusion.
The conclusion is that geographic origin and brand prestige do not appear in the benchmark results. Capability does. Price does. Humans are, when sufficiently motivated by finance, capable of noticing this.
What happens next
Databricks says it plans to route coding work more intelligently across models by complexity tier, running the cheapest capable model for each task. Co-founder Matei Zaharia has described this as the evidence showing it is time to deploy these as daily drivers.
The Western AI labs, who spent considerable resources establishing themselves as the expensive default, are invited to compete. The benchmark does not ask where you are from. It only asks if you pass.