Hey @jaketobxd! Your follow-up is correct. The extra API usage isn’t coming from Auto or Grok 4.5 themselves.
Your included usage is split into two separate pools of very different sizes:
- First-party models: Auto, Composer 2.5, and Grok 4.5 (including the non-fast / Medium effort variants). Much larger allowance.
- API: third-party models (Claude, GPT, Gemini) at their API price. Much smaller allowance.
Building on @rsb’s point: both count toward your included usage, but they’re two separate pools with separate limits (not one shared bucket), billed at different rates. Since the budgets are so different, the two percentages move at very different speeds, which is why you can sit at ~1% first-party while API is already ~13%. The thread rsb linked covers that math well.
So Auto and Grok 4.5 (any effort or fast setting) are correctly billing to the first-party pool. What’s actually draining your API pool is subagents: while working in an Auto or Grok session, the agent’s Task tool sometimes spins up a subagent on a premium model like Claude Opus or Sonnet that you didn’t pick, and those calls bill to the API pool.
This is a known issue, and the team is working on improving how subagents choose their model. Until that lands, the most reliable workaround is to watch the subagent cards in the transcript and cancel any running on a premium model you didn’t ask for. Setting your subagent model to a first-party model (Composer 2.5 or Grok 4.5) in your model settings can help too, where it applies.
You’re still inside your included API usage (the ~13% you mentioned), so there’s no on-demand charge here. Let me know if anything still looks off!