Claude telemetry: 4% of sessions ran up 65% of the bill

Two sessions ran up 32% of a ten-person team’s spend. The team kept both open and resumed them every day for over a week. The figure comes from 310,009 OpenTelemetry events logged over 30 days in Claude Code and Cowork.

Most of the money paid for memory. Cache reads made up about 68% of spend. That’s the agent re-reading its own conversation, and in those two sessions every call paid for ~500k tokens of cached context.

On security, Claude printed credentials outright 66 times. Each value stayed in context until compaction, so 9% of model calls sent a credential to Anthropic in plaintext. And 14% of turns read untrusted external data, with MCP servers as the biggest source.

The data covers one anonymized enterprise team, and the write-up doesn’t say what those two sessions were working on.