Databricks reports 60% higher coding spend after rolling out Astra to 3,500 engineers

Databricks publicly disclosed that rolling out GPT-6 Astra to approximately 3,500 engineers increased total coding spend by roughly 60% compared to prior top-end models like Opus 5 and Sol 5.6. The company attributed the higher spend to Astra’s superior performance on high-complexity system design and long-horizon tasks, which justified its use despite the cost increase.

In an internal note, Databricks engineer Patrick Wendell reported that Astra ‘unambiguously’ outperformed prior models on complex tasks, particularly those involving high-level system design. However, the note also emphasized that Astra did not materially improve performance on medium or low-complexity coding tasks, limiting its universal applicability within workflows.

To manage the cost impact, Databricks created a dedicated sub-budget for Astra, encouraging selective use of the model within engineering teams. This approach reflects a broader industry trend where premium long-horizon models are reserved for tasks where their capabilities provide outsized value, even at higher operational costs.

The company’s experience with Astra is part of a growing pattern in which benchmarks and internal evaluations increasingly position Astra at or near the top of overall capability rankings. For example, EpochAIResearch noted that Astra now leads their Epoch Capabilities Index, including setting a new Math-ECI record, while Anthropic’s Claude Fable 5.1 remains a top performer in software engineering benchmarks.

Arena’s comparative data further illustrates the cost-performance trade-off: Astra Max was reported at +$11.7% / $3.94 per task versus Sol xHigh at +$7.0% / $1.03, while Fable 5.1 Max was at +$13.7% / $4.40 versus Opus 5 High at +$10.2% / $2.07. On web development-specific arena data, Astra ranked first overall, though Fable 5.1 remained preferred in some head-to-head comparisons.