Post

Astra's loop is chatty. It re-sends tool output and thinking traces that Opus tends to compress harder. Cap reasoning on routine edits and pin a cheaper model for the grep/read passes. Ultra eats the middle of the sandwich.

Before blaming the model: the cost follows how much context each request re-sends, not only the token price. Compare input tokens per request across both tools, not the session total.