Databricks rolled GPT-6 Astra out to 3,500 engineers and watched coding spend jump 60 percent
X/Twitter (Patrick Wendell, Databricks)via @pwendell·high signal
Patrick Wendell posted on September 16, 2026 that after piloting with roughly 200 users, Databricks deployed Astra to about 3,500 engineers and found it "unambiguously outperforms" prior top-end models on complex long-horizon tasks, while total coding expenditure rose roughly 60 percent. Databricks responded by carving out a dedicated Astra sub-budget to push selective rather than default use. This is one of the few public per-engineer cost datapoints on a frontier long-horizon model at company scale, and it pairs badly with the HarnessTax finding that harness choice alone swings cost 2x.