If you’ve been running heavy coding, research, or agentic tasks lately and noticed Claude Opus 5 eating your limits faster than Sol, you’re not imagining things.
The short version: Claude is really good, but it's also an expensive over-thinker.
For harder tasks, Claude can use extended thinking, meaning it spends a bunch of tokens reasoning before you ever see the answer. Those thinking tokens still count.
Then tool use makes it worse. In long chats, every new step can drag a lot of prior context back into the next request. So if the model takes ten tiny careful steps instead of three bigger ones, your usage can balloon fast.
Sol tends to be more economical on a lot of these long-horizon workflows because it often gets to the result with fewer tokens and fewer loops.
Quick fixes:
>> Cap the thinking budget when you can.
>> Ask for “final result only” when you don’t need progress narration.
>> Periodically summarize or restart long chats instead of letting the full transcript trail behind every request.
>> Claude isn’t “bad” here. It’s just thorough in a way your usage meter absolutely notices.
