Feature: Consumer overage calibration mode

[nhaener]

Problem

The current API-rate estimates in Tide Pools are based on direct Anthropic/OpenAI API pricing. For users on the Claude.ai Max consumer plan, actual overage charges diverge significantly from these estimates — possibly by a large factor.

Root causes likely include:

• Consumer plan may not apply cache read/write discounts
• Blended per-model overage rate that differs from direct API pricing
• Possible minimum billing granularity or overhead terms

Example: Our setup showed ~$5.55 estimated via Tide Pools but ~$30 in actual claude.ai overage charges over the same period (~5-6x divergence).

Proposal: Calibration Mode

Add an optional calibration flow where users can input their actual billing charges and fit effective per-model rates:

1. User inputs actual claude.ai overage charge for a given period
2. Tide Pools compares against its token-count-based estimate for that period
3. Fits effective per-model rate coefficients (+ optional global overhead term)
4. Stores calibration factors locally
5. Outputs both Estimated charge (raw API rates) and Calibrated charge (fitted rates) with an error % confidence band

Output Example

By Model:
  claude-sonnet-4-6: 6.4M tok · $4.88 est · $24.40 calibrated (5.0x)
  claude-haiku-4-5:  815K tok · $0.44 est · $1.76 calibrated (4.0x)

Total: $5.32 est · $26.16 calibrated (±15%)

Notes

• Calibration factors should recalibrate weekly as more data accumulates
• Should degrade gracefully if no calibration data exists (show raw estimates with a disclaimer)
• Could support multiple billing profiles (consumer Max vs direct API vs other plans)

cc @clawSean — discussed in Dev Team chat, Sean proposed the regression model approach