Throttle

Throttle Detection answers a frustrating question: how much of your day is lost to waiting, and is it something you can fix or a Claude-side problem? It watches for the moments when the AI makes you wait, adds up the time, and — the useful part — separates the waits you can do something about from the ones you can’t. It shows your own activity.

Reading the headline

Three numbers sit at the top:

• Throttle rate — the share of your prompts that ran into throttling. A rising rate means you’re hitting limits more often.
• Wait time — the hours you lost to it in the range. This is deliberately conservative: a normal-but-slow response or a coffee break isn’t counted, only real throttling.
• Trend — this half of the range against the previous half, so you can tell whether it’s getting better or worse.

Below them, the wait is split into Fixable and Platform, and a Top cause line names the detector behind most of it. That split is the whole point: Fixable time responds to how you pace and phrase your work, so it’s worth chasing; Platform time is Claude-side and out of your hands.

Where the time went

• Wait time over time — a stacked bar chart, one bar per period, colored by the detector that caused each stretch of wait, so a bad day stands out at a glance.
• Throttle events — a table of the events that cost you the most, with When, Detector, Details, Hours Lost, and Session. If there were Claude-side outages in the range, a banner up top notes how many hours they affected and whether they matched a known incident.

The four detectors

Each wait is attributed to one of four causes:

• Rate limits — you hit a 429/529 and had to wait. Fixable — spreading work out or trimming parallel requests helps.
• Latency spikes — a response came back unusually slow. Fixable.
• Prompt retries — a request had to be retried. Fixable.
• Service outages — a Claude-side incident. Platform — nothing you do changes it, so it’s kept out of your throttle rate.

Getting time back

Click any event for the Event Detail panel: what the detector saw, the estimated delay and hours lost, and when it happened. For latency spikes and prompt retries it also shows a How to fix list with do/don’t examples — the concrete change that would have avoided the wait. Start with your biggest Fixable cause and work down.