Drop in the usage export from OpenAI, Anthropic or OpenRouter. In two minutes you'll see where your AI bill goes and what you pay for nothing. Before you switch to anything cheaper, we test it on your real requests.
Free. Your file is read in your browser and never reaches us.
On your phone? Email yourself the link for your laptop.
Gosset measures each of these in your own traffic and puts a monthly dollar figure on it.
Your system prompt goes out with every request, and you may be paying full price for it every single time.
A timeout fires, your code retries, and you pay for both calls.
Reasoning models think before they answer, and every hidden token is billed, even when the answer is a single word.
Nightly reports and background jobs that don't need an instant answer, billed at the instant price.
When a model your app calls is retired, its replacement can cost more or answer differently. We test the replacement on your requests first.
The model you picked at launch is rarely the cheapest one that still works today. We find the one that is, and show you which answers would change.
This page reads your file inside your browser, so it never reaches us. If you want to check, turn off your wifi after the page loads. It still works.
You see your numbers before we ask for anything.
Spend by model, the share of your input served from cache, retries you paid for, and any model with a shutdown date.
Free, try it now →95% of input paid at full price
Paste one line into Claude Code or Cursor. Your agent sends us a sample of real requests with personal details masked. We test the replacement and cheaper models and show you every answer that got worse.
$149 one time, per appToday's model quoted the 14-day refund window from message 3. The cheaper model missed it.
We test each week's new models on your saved requests. You only hear from us when a switch would save money or a shutdown is close.
From $29 a monthThe change is ready to paste into your coding agent. Nothing else needs your attention this week.
Anything else: contact us.
On the sample bill, $175 to $415 a month on a $1,089 bill. Yours depends on your caching, retries, batch use and models, so we don't quote a percentage. The free X-ray shows your own numbers in two minutes: what you spend on each model, how much input you pay full price for, and what is going to models that are shutting down. The break test then measures what a cheaper model would save on your real requests, and what it would break.
No. The X-ray is plain JavaScript running in your browser. The file is read on your device and never sent to us or anyone else. The page loads no analytics or outside scripts.
OpenAI and Anthropic exports contain token counts but no prices, so we multiply your tokens by each model's published list price, including cached and batch rates. If you have negotiated prices or credits, your real bill will be lower. OpenRouter exports include the actual cost of each request, which we use as is.
From OpenAI's and Anthropic's published deprecation pages. We check them every week and show the date we last checked under your results.
Your coding agent masks emails, phone numbers, card numbers, addresses and names before anything leaves your machine. We use the masked requests only to run your test, and delete them after 14 days unless you join a weekly plan, which keeps them as your test set until you cancel.
Sometimes, not always. Even when a new model scores better on average, studies of real migrations find that 4% to 15% of individual requests get worse. That is why we test on your own requests and list each answer that changes, instead of quoting a benchmark.
William Sealy Gosset was a Guinness brewer who invented a statistical test for telling whether a cheaper batch is as good as the original, from only a few samples. That is what we do with AI models. Read the story.