Your LLM bill is 5x what it should be. We metered it.

We metered 5 of our own agentic sessions — 787.7M input tokens, 6.35M output. The ratio is 124:1. For every dollar of answers, $124 of context got re-sent. Anyone selling you shorter answers is optimizing the crumb.

Then the good news: 94% of that input was cache-served (288.8M of 304.3M on our longest session). At October 2026 rates on Anthropic, OpenAI, Google, xAI, and DeepSeek alike, a 124:1 session at 94% cache pays roughly a fifth of unprofiled cost. One 304M-token marathon: about $116 instead of about $631.

The catch is that the fifth isn't automatic. It holds only while prefixes stay stable, polls stay rare, reads stay bounded, and re-runs stay earned. Drift on any of those and the cache share collapses — which is exactly what PSI watches: every model call tagged with hours, memories, dollars, tokens, carbon, and time, with the three biggest leaks named first.

PSI Meter is $19/month. Org is $79. A 10M-token-a-month user saves about $17.60 of that $20.80 Sonnet-class bill — the meter pays for itself eleven times over before lunch. Your first audit is free: send us session logs, we hand back your ratio, your cache share, and your three leaks.

Measured numbers, dated, reproducible from our own logs. No 92% anything — if someone quotes you one without a meter behind it, ask to see the meter.