Based on an internal benchmark at Databricks evaluating coding agents on actual tasks from its multi-million line codebase covering Python, Go, Typescript, and Scala, the analysis found that the Pareto frontier for coding tasks includes models from OpenAI, Anthropic, and open source, meaning only a mix of tools provides frontier performance. Open models, and GLM 5.2 in particular, can handle the highest level of task difficulty. The token price of a model is a poor indicator of actual costs, as larger models can be more token efficient with lower overall costs. The harness a model is called from dramatically impacts cost and quality, with simple harnesses like Pi performing best on Databricks' workloads. Results showed clear clustering into three capability tiers: the most intelligent models are very effective but expensive, while medium and lower intelligence models are highly effective for common tasks and significantly cheaper. Based on this analysis, Databricks determined it should push more work to the Haiku and GPT 5.4 M models.