External variables (as of February 2026)

  • Input: $1/MTok

  • Output: $5/MTok

  • Tokens per second: ~110 TPS

Assumptions

  • Typical ratio of input to output tokens: 5:1

  • No caching, using extended thinking

  • 5% of time is downtime: constructing prompts, waiting for tool calls, time to first token, etc.

Steps

  • 110 TPS x 3600 sec/hr = ~0.4 MTok/hr output = $2/hr

  • 5:1 ratio means ~2 MTok/hr input = $2/hr

  • Total of $4/hr with input and output tokens combined

  • ~$3.80/hr assuming only 5% downtime

Result

  • Running Haiku 4.5 continuously costs ~$3.8/hr

Interpretations

  • If you are running an agent 12 hr/day, 30 days/month, it comes out to ~$1.4k per month

  • If you are limited to $100/month, you can pay for ~50 min/day

Adjustments for Continuous Opus 4.6 usage