2.1B tokens in a single day for $16. V4.1 Flash 1017 is no joke.

I've spent the last week running v4.1 Flash 1017 on one of my real projects: ~6B tokens, 25K+ requests, $46 total. My peak day alone was 2.1B tokens for $16 (screenshots attached). Most of that was cache hits. TL;DR: Not Sol 6.1 level and nowhere near Opus 5.5, but if you delegate to it properly as an agent, it gets real work done. Where it shines Well-specified tasks with a clear scope Running and iterating on tests Agentic work when you break things down for it Where it slips Still hallucinates sometimes, though noticeably less than before Fills gaps between tasks with its own guesses instead of flagging them As an orchestrator, it occasionally skips some playbook instructions Overall I've been loving it. Give it a tight spec and it just goes. One thing I haven't done yet is audit the code it wrote for vulnerabilities and run a proper security analysis. That's next on my list, and I'll post the results here if anyone's interested. Let me know in the comments. Curious what yall are seeing: are you running 1017 as an orchestrator, or keeping it to worker/sub-agent tasks?