Find out what your AI habit actually costs before the bill does

The bill that surprises people almost never comes from the model they expected. It comes from the workflow nobody was watching -- an agent looping through the same file five times, a chatbot resending an entire conversation's history on every single turn, a batch job that quietly ran overnight. The per-million-token price on a vendor's homepage answers a different question than the one that actually determines your bill. Here's the 15-minute audit that finds where your money is actually going, and the one setting that stops it from happening again.

Why the sticker price is the wrong number to stare at

Every major lab prices output tokens -- the words a model generates -- at roughly five times what it charges for input tokens, the words you send it. That ratio holds almost exactly across three otherwise very differently priced models: Anthropic's Claude Fable 5.1 ($10 input / $50 output per million tokens), OpenAI's GPT-5.6 Sol ($4 input / $20 output, after an August 21 price cut), and Google's Gemini 3.8 Flash ($0.75 input / $3.75 output). A workflow that makes a model write a lot -- long reports, generated code, multi-step reasoning traces -- costs far more per exchange than one where you paste in a document and ask a short question about it, even at the identical sticker price.

That last row is where agentic workflows actually get expensive or cheap. Cache reads -- reusing a chunk of context you already sent, instead of paying full input price to resend it -- cost a small fraction of a fresh input token. Anthropic cut its own cached-input price by 75% -- to $0.25 per million tokens -- when Fable 5.1 shipped (OpenAI's equivalent cached rate on GPT-5.6 Sol is $0.40 per million), specifically because agentic, multi-turn sessions were re-sending the same growing conversation history over and over, and that repetition was where the real cost sat. If your workflow doesn't structure its prompts to actually hit the cache -- a stable system prompt and history prefix, with only the new part changing -- you're paying full input price for content the model has already seen ten times in the same session.

Here's what that looks like in practice. Say you point an agent at a mid-sized codebase and ask it to fix a failing test suite. A naive setup resends the entire relevant file tree as part of the prompt on every single tool call the agent makes -- read a file, resend everything; run a test, resend everything; edit a line, resend everything again. Forty tool calls into a real debugging session, you've paid full input price for the same unchanged files roughly forty times over, even though maybe three of those calls actually needed fresh information. None of that shows up as a warning anywhere. It shows up as a number on the dashboard that's ten times bigger than the task felt like it should have cost -- which is exactly why the audit below starts at the dashboard, not at the prompt.

One more thing worth saying plainly before the steps: a flat-rate consumer subscription -- ChatGPT Plus, Claude Pro, Gemini Advanced -- doesn't expose any of this, because you're not paying per token. If that's the only way you use these tools, this audit doesn't apply to you yet, and that's a feature of the plan, not a gap in this guide. It becomes relevant the moment you, or a tool you've connected, starts calling a model through its API or building an agent on top of one -- which is precisely when a runaway loop stops being capped by a monthly subscription price and starts being capped only by whatever budget alert you did or didn't set.

The 15-minute audit

That fourth step -- context resent from scratch -- is worth a concrete example, because it's the one that fools people who already know about token pricing in the abstract.

Run that estimate before you scale a workflow up, not after -- it costs one exchange and it's the cheapest insurance against the fifth pitfall below: finding out at the invoice instead of before you hit run.

Whichever branch applies, the fix is almost always smaller than the audit made it feel -- a length cap, a reordered prompt, or a retry limit, not a rewrite.

The four failure modes above share one root cause: treating the sticker price, or a rough estimate, as if it were the actual bill. The ledger below is the last piece -- what that sticker price does and doesn't include, stated plainly.

None of this requires enterprise FinOps software or a dedicated cost-management hire -- it's fifteen minutes with a dashboard you already have access to. What it buys you is the difference between finding a runaway workflow the day it starts and finding it on next month's invoice. The sticker price on the homepage was never going to tell you that; only your own usage was.