Databricks Shares Five Observations After Full Deployment of GPT-6 Astra

Databricks co-founder Patrick Wendell shared five observations from deploying GPT-6 Astra across the company. Key findings: capability gains are concentrated in complex tasks, overall coding expenditure rose 60%, and the company uses a gateway with queue experiments for rollout and budget-based routing to steer model selection.
Key insights
Capability Gains Concentrated in Complex Tasks
Astra clearly outperforms previous flagship models (Opus 5, Sol 5.6) on high-complexity tasks like system design and long-horizon tasks, but shows no perceptible improvement on medium- or low-complexity tasks. The author suggests a 'saturation hypothesis': existing models already perform simple tasks near perfectly, leaving no room for improvement. This implies that upgrading models may have limited ROI for teams with mostly simple tasks.
Cost Increase of 60% and Budget-Based Routing
Overall coding expenditure rose about 60% after Astra deployment, likely because better models encourage more usage and complex tasks that were previously avoided. Databricks treats this as a signal for governance, not a success metric. They implemented budget-based routing: Astra has a separate sub-budget, guiding engineers to use it for complex tasks and cheaper models for daily work, turning model routing into a resource allocation problem.
Rollout Methodology: Gateway and Queue Experiments
Databricks used a staged rollout: 200-person pilot to gather quality and cost signals, then expanded to 3,500 people. All traffic went through the Unity Gateway, which served as a control plane for model evaluation, canary releases, and budget management. This demonstrates an emerging industry standard where new models are introduced experimentally like feature flags.