2026-07-10
AI Tokenomics: what tokens actually cost
The foundation: what tokens are, input vs output pricing, cached tokens, long context, and why agents multiply everything.
LEARN / THE SERIOUS PART
AI costs, explained
The editorial side of the TokenBurn Index: how AI billing actually works, where the money goes, and which optimizations are real. Everything here follows the house rules — numbers you can check, assumptions stated, no environmental salvation promised.
2026-07-10
How much does ChatGPT actually cost per prompt?
Subscription math vs API pricing, the conversation-history multiplier nobody budgets, and when each way of paying wins.
2026-07-10
Prompt caching: the 90% discount most teams never claim
How cached input tokens cut LLM bills by up to ~90%, the write-premium fine print, and the silent mistakes that break your cache.
2026-07-10
Why your AI agent costs 10–40x more than a chat
The anatomy of agent costs: iteration loops, re-sent context and invisible reasoning — with worked numbers and the levers that reduce the bill.
2026-08-04
Why output tokens cost more than input
Output is priced above input, hides reasoning tokens you never read, and balloons when an agent rewrites unchanged code. How to cut generated-token spend.
2026-08-04
Cut AI costs with model routing
Route routine tasks to a cheaper model, keep the flagship for the hard cases, and cut the bill — with worked numbers and the escalation math that decides it.
2026-08-12
The cache minimum: why your prompt cache silently does nothing
Teardown #1: every Anthropic model publishes a minimum prompt length for caching. Below it nothing caches, at full price, with no warning — and one token decides it.
2026-08-13
What a Cursor session costs, by model
Teardown #2: Cursor bills on-demand usage at raw API rates by its own docs, so our own priced coding-agent session tells you what the $20 pool buys — and the spread across models is over 100x.
2026-08-14
Four months of a coding agent, measured
Teardown #5, field notes: 18,700 agent messages and 2.35 billion tokens of context, measured. For every token sent fresh, 35 came from cache — and that ratio moves the bill more than the model does.
2026-08-13
The tokenizer tax: a price rise with no price change
Teardown #4: Anthropic's newer models turn the same text into ~30% more tokens. The rate never moved, so no price tracker sees it — and a sticker-price comparison understates the real gap by that much. Including in our own calculator.
2026-08-13
The Sonnet 5 price cliff that never happened
Teardown #3, corrected: the scheduled 50% hike on Sonnet 5 was cancelled. What it would have cost, and how our own price note asserted it for weeks after it stopped being true — with every automated check passing green.
2026-08-04
How to read your AI usage export
Read OpenAI and Anthropic usage CSVs: normalize the token columns, spot oversized prompts and weak caching, and reconcile your estimate with the real invoice.
Reading is cheap. Measuring is free. Paste your prompt into the Burnmeter and see what your AI habits actually cost.
MEASURE MY BURN →