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

Databricks benchmarked GLM 5.2 on its own multi-million-line codebase and found the Chinese open-source model statistically tied with Anthropic's Opus 4.8 at lower cost. The company now plans to make it a daily workhorse for its developers.

GLM 5.2 hit the top performance cluster at $1.28 per task versus $1.94 for Opus. "The evidence shows it's time to start deploying these as daily drivers for coding," write the authors of the blog post, including Databricks co-founder Matei Zaharia. Developer feedback from internal pilots backed up the results, and the company says it's already working on running GLM at peak performance.

Databricks isn't alone. 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.

No single lab dominates across three performance tiers

Overall, the tested models and configs fell into three clusters, according to Databricks. The top group, with an 82 to 90 percent pass rate, includes Opus 4.8, GLM 5.2, and GPT 5.5 in certain configs. A middle group at 71 to 82 percent includes Sonnet 4.6, Sonnet 5, and GPT 5.4, among others. The bottom tier at 51 to 60 percent holds GPT 5.4-mini and Haiku 4.5.

An analysis through Unity AI Gateway found that 61 percent of coding tasks from Databricks engineers are medium complexity, about 19 percent low, and only 12 percent high. The most expensive models had been the default. Now the company plans to route more work to cheaper tiers based on task complexity.

The Pareto frontier, the best quality-to-cost ratio, is shaped by models from three providers: OpenAI, Anthropic, and open source. Only a mix delivers frontier-level performance, Databricks says.

Databricks also points out that token price and actual task cost aren't the same. Token efficiency matters just as much, like fuel economy in a car, and varies widely by software environment. In one test, the Pi harness sent about three times less context than Claude Code. For Opus 4.8 at "high effort," Pi was 2.08x cheaper at comparable quality (85 versus 87 percent). GPT 5.5 showed a similar pattern: Codex used 1,235,000 tokens versus 665,000 for Pi.

Real tasks instead of public datasets

Databricks built its own benchmark from real pull requests rather than relying on public alternatives like SWE-Bench. Solutions leak into training data over time, and the tasks don't match a stack spanning more than ten languages, including Python, Go, TypeScript, Scala, and Rust. OpenAI recently warned against SWE-Bench-Pro for similar reasons.

Each task had to be recent, human-written, paired with high-quality tests, and representative of the full stack. All were reviewed by hand, with tests partly rewritten to allow alternative implementations. Scoring relied solely on passing tests, not an LLM judge, which Databricks says tends to reward answers that sound good rather than ones that are correct.

The team also hit a cheating problem: models searched the Git history for the correct solution instead of working it out. Databricks fixed this by truncating the entire Git history for each run.