A $115-an-hour agent loop can drain a $500 monthly hard cap in about four hours

The $115 figure is a scenario with a condition attached. The agent is in a retry loop on Claude Opus 5.5, and every turn re-sends a 150K-token context uncached ^[24]. The passage does not include the per-token price or the turn rate behind the number. For it to transfer to your workload, your loop has to send the full context uncached on every turn. A workload that caches its context is outside it.

When a soft alert's threshold is crossed, someone gets an email and requests still go through. Crossing a hard cap's threshold gets an error back instead ^[14]. In OpenAI's guide, spend alerts are described as not enforcing a cap ^[18]. An unattended agent produces spend at machine speed, and the alert goes to a human who is asleep ^[15]. Simon Willison made the same argument on 3 October 2026, according to the post: a soft cap is not enough, and most people would rather get an error than an unexpected $10,000 bill ^[16].

Take the post's example, a monthly $500 cap. The post says one bad night can eat a large share of it before anyone looks ^[9]. At $115 an hour the loop reaches $500 in about 4.3 hours ^[22]. Lag adds to that. Google says billing data for Gemini project caps can trail by about 10 minutes ^[3], which is roughly $19 of extra spend at the same rate ^[23]. Google also says actual costs on Google Cloud can take more than 24 hours to process ^[4].

The scopes differ. OpenAI enforces at organization and project level ^[19]. Google Cloud's Spend Caps cover one project and one eligible service, and requests already in flight finish and accrue charges ^[13]. Vercel budgets apply at team, project, API key and team member scope ^[6]. A team on several providers sets a differently shaped limit in each.

Anthropic's row has the widest gap. The post says Anthropic documents the limit but not the failure behaviour ^[21]. Google's tier cap also has no documented error code ^[20]. We think that is the first thing to check, because a retry loop can only stop on an error it recognises. OpenAI's cap returns a named code for each scope ^[2], so a harness can match it. Where the code is undocumented, the harness is guessing.

AWS is the least settled row. The post calls AWS project spend limits a limited release and could not verify Bedrock-specific behaviour on any official AWS page ^[25]. It relays Willison quoting AWS that the project is paused for the month, and a Hacker News commenter's report of a deletion warning after 90 days, and says to treat both as unverified ^[26].

By the post's account, every cell in its table comes from vendor documentation or a source named in the row ^[17]. The table therefore records what vendors say a cap does, and says little about how one behaves under a loop. In our view a vendor cap is the right ceiling for an unattended agent. It needs a faster control beneath it, because a monthly limit that trips at hour four has already used the whole month's allowance ^[22].