Databricks rolls out Astra to all ~3500 engineers, coding spend jumps 60%

pwendell · x · 2026-09-17

Paul Wendell of Databricks shared takeaways from rolling out Astra to all 3500 engineers:

Capability: Astra clearly outperforms their previous top models (Opus 5.0, Sol 5.6) on highly complex tasks, especially high-level system design and long-horizon work. On medium/low-complexity coding tasks the gains are unclear — those tasks are likely saturated by existing models.
Usage: Engineers given Astra increased overall coding spend by 60% vs baseline.
Method: They piloted with 200 users to gather quality and cost signal, using Unity AI Gateway for cohort-based experiments on all new models.
Cost control: Engineers get a separate Astra sub-budget to encourage selective use on complex tasks while preferring cheaper models for everyday work, with free mix-and-match within an overall budget.
No robust Astra-vs-Fable comparison yet, as Fable hasn't been widely rolled out due to data retention policies.

More from coding & agent