The Pulse
Databricks Expands Astra to Every Engineer
Databricks has rolled out Astra to roughly 3,500 engineers after a pilot with about 200 users, while using budget controls to guide model selection and spending.
AI.info Team ·
Databricks Expands Astra to Every Engineer
Databricks has rolled out Astra to every engineer at the company, covering roughly 3,500 users after piloting the model with around 200 people. Patrick Wendell said the pilot was used to gather signals on both quality and cost.
“Today we rolled out Astra to every engineer at Databricks (N=~3500).”
Patrick Wendell, Databricks co-founder and vice president of engineering
Wendell said Astra outperformed Databricks’ previous highest-end models, including Opus 5 and Sol 5.6, on highly complex tasks. He highlighted work related to high-level system design and long-range tasks as areas where the difference was especially clear.
The advantage was not consistent across all coding work. Wendell said it was not clear that Astra meaningfully improved medium- and low-complexity coding tasks compared with earlier models. Databricks suspects those tasks are mostly saturated, meaning existing models already execute them effectively.
Spending Rose Among Astra Users
Engineers given Astra increased overall coding spend by around 60% compared with baseline. The figure describes coding-related usage costs for that group rather than headcount or engineering time.
Unity Gateway Sets a Budget Boundary
Databricks uses Unity Gateway to run cohort-based experiments for new models. The company gives engineers a sub-budget specific to Astra, encouraging selective use of the model on complex tasks while allowing lower-cost models for everyday work.
Engineers can mix and match tools and models within their overall budget envelope. Wendell said Databricks also allows budgets to be increased through various mechanisms, with budgets defined in Unity Gateway and regularly revisited.
That arrangement gives the company a way to make Astra broadly available without making it the default choice for every coding request. Engineers can use the model where its additional capability is most valuable while relying on less expensive options for work that existing systems already handle effectively.
A Narrow Advantage on Difficult Engineering Tasks
Databricks’ findings place Astra ahead of its previous highest-end options on tasks involving system design and long-range work. Those assignments can require an engineer or coding agent to understand a large repository, weigh competing architectural approaches, change multiple components and verify that the pieces work together.
The result does not establish that Astra is better for every engineering workflow. It points to a more limited conclusion: the model’s advantage is most apparent when a task is complex enough for extended reasoning and sustained work to matter.
For routine development, the company’s evaluation found less meaningful separation between Astra and earlier models. That leaves Databricks with a cost-allocation question as much as a model-selection question: how to reserve the most expensive system for work that benefits from it while keeping ordinary coding within a broader budget.
What the Rollout Shows
Databricks’ deployment is a large internal test of Astra at engineering-team scale. The company moved from a pilot involving around 200 users to access for roughly 3,500 engineers, then used budget controls to encourage different models for different kinds of work.
The reported outcome is mixed but specific. Astra outperformed Opus 5 and Sol 5.6 on highly complex tasks, showed no clear meaningful improvement on medium- and low-complexity coding, and was associated with an overall coding-spend increase of around 60% among engineers given Astra compared with baseline.
Wendell also noted that Databricks does not have robust Astra-versus-Fable comparisons because Fable has not yet been rolled out widely due to data-retention policies.