Production AI economics guides

Measure production AI costs before making a change.

Practical methods for finding the real unit cost of an AI workload, comparing alternatives, and verifying whether a change improved the business outcome without crossing its operating constraints.

01 / Foundations

Start with a defensible unit metric.

Cost per token is useful for checking a provider bill. Cost per successful outcome is more useful for comparing production decisions.

Guide · 9 minute read

How to calculate AI cost per successful outcome

Choose a denominator, include retries and human review, separate operating cost from implementation expense, and work through an illustrative comparison.

Read the guide Guide · 11 minute read

Production LLM cost-optimization checklist

Trace calls, context, output, retries, caching, batching, and routing in the order needed to turn a rising bill into a testable change.

Use the checklist Guide · 12 minute read

API vs. self-hosted AI: a TCO guide

Compare provider usage with production capacity, idle reserve, engineering labor, reliability, implementation, and cost per successful outcome.

Compare the options Guide · 8 minute read

Reduce AI API costs without losing quality

Sort cost cuts into ones that need no evaluation and ones that require testing against a written quality floor before they ship.

Sort the tradeoffs Guide · 8 minute read

LLM retry costs: the hidden multiplier

Reconcile the two definitions of retry rate, price nested retries and workflow restarts, and log the reason for every extra attempt.

Measure retries Guide · 8 minute read

Prompt caching cost savings: a break-even guide

Solve for the reuse count where a cached prefix's write premium pays for itself, and the traffic conditions that break the savings.

Find the break-even Pricing snapshot · September 2026

Claude vs. GPT-6 API pricing

Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 4.5 against GPT-6 Astra, GPT-6.1 Sol, and Luna, with a cached-workload comparison.

Compare the rates Measured data · 91,086 agent calls

Where AI agent tokens go

Claude Code and Codex logs, counted and priced: cache is 82% of cost, context grows 7x per session, and 10% of sessions drive 80% of spend.

See the data

Before a private diagnostic

Use public methods to prepare your data.

Each guide explains a reusable method, states its assumptions, and uses illustrative examples. Verify modeled savings against production data.

Use the free tools