Yegge shuts Gas Town as Databricks flags 60% Astra cost

The two reports land together because they test the same assumption from opposite directions. Gas Town was the most prominent public example of the 'orchestrator' pattern — a tool whose purpose was to coordinate coding agents at scale — and its author's admission that he never successfully built anything with it undercuts the case for that layer of tooling. Databricks' rollout tests the other assumption: that a premium long-horizon model will pay for itself when handed to every engineer. Astra did outperform Opus 5 and Sol 5.6 on the hardest work, but the ~60% spend increase arrived alongside an explicit note that medium- and low-complexity coding may not improve much, which is where most day-to-day engineering time sits. That mismatch between where the model wins and where the volume is likely explains the sub-budget response.

The debate over what these two data points mean will likely hinge on how much of the added cost converts into shipped work rather than how the models score on benchmarks. Teams that adopt premium models per-seat will probably face the same conversation Databricks preempted with a sub-budget, while teams that invested in orchestrator layers may re-examine whether the reliability of task completion justifies the coordination overhead. For vendors, the relevant question is whether selective, high-complexity use — rather than broad seat distribution — becomes the default enterprise go-to-market motion.