Inference cost calculator

Frontier models against open-weight alternatives on hosted inference, annualised and multiplied across a portfolio. Every rate carries an as-of date.

Inputs

Start from a profileCurrent model

Sizing input

Monthly spend, USDWorkload typeOpen-weight alternative

Multiply across portfolio×

The share card is a 1200×630 PNG of this result — sized for a slide, an email, or LinkedIn. Rates as of 2026-07-01.

Annualised comparison

Frontier · Claude Sonnet

$480,000

annualised · $40,000/mo

Open-weight · DeepSeek (hosted)

$76,000

annualised · $6,333/mo

Savings delta, annualised− $404,000/yr

Across 25 portfolio companies− $10,100,000/yr

Assumptions

Token volume (derived)6,061M tokens/mo

Workload mix70% input / 30% output

Claude Sonnet rate$3.00 in / $15.00 out per 1M · as of 2026-07-01

DeepSeek (hosted) rate$0.55 in / $2.20 out per 1M · as of 2026-07-01

Assumes eval-gated routing: the cheaper model must clear a per-workload quality floor before traffic moves. Where evals fail, the spread narrows — this is an estimate, not a quote.

This delta is an estimate until it's verified against your own traffic.

A shadow-mode audit mirrors a week of real traffic and returns a verified per-use-case number — nothing changes in production.

Get your number verified → o10.ioshadow-mode audit · a week of traffic · no production change