Three years ago, "AI cost" wasn't a line item. A handful of seats on ChatGPT Plus, a Cursor subscription, maybe a hobby OpenAI key — total damage under $100. Nobody looked.

Now your team's collective Anthropic invoice is $4,200 for the month. Cursor charged $20 a seat, the API bill is $1,800, ChatGPT is another $600. The CTO is asking questions. The CFO has more questions. Nobody on the team can answer them — because nobody can see where the money actually went.

This is the spend visibility problem. It's the exact same one cloud teams hit around 2014 when AWS bills crossed $10k/month and nobody knew whether the spike was a misconfigured Lambda, a forgotten EBS volume, or just normal growth. The answer back then was tools like Vantage, CloudHealth, and AWS Cost Explorer. The answer for AI spend is per-prompt token visibility.

The invoice is a lagging, lossy signal.

Provider invoices give you one number per month: total spend. They sometimes split by model. They never split by who, by project, by prompt, or by retry.

That means by the time you see a $200 over-budget month, you have:

Without that detail, the only response to a high bill is panic-cutting. Cut seats. Switch the team to a smaller model. Ban Cursor on Saturdays. All blunt. All bad.

Tokens are the unit. Not dollars.

AI pricing is denominated in tokens — input tokens, output tokens, cached tokens — each with its own rate that varies by model. Sonnet input costs roughly 5× a Haiku input. Opus output costs 5× Sonnet output. When you optimize at the dollar level you can't see any of that; when you optimize at the token level the levers become obvious.

A real example: a 12k-token prompt asking Opus to summarize a meeting transcript costs about $0.18. The same prompt against Haiku costs about $0.01. If the summary is going into a Slack channel for a quick stand-up — Haiku is fine. If you keep using Opus on autopilot because "it's smarter", that's $0.17 per summary, every day, forever, for an output nobody can tell the difference on.

You can't make that call without seeing the tokens. The dollar number is too coarse.

~30%

Of typical AI spend goes to repeated or retried prompts — the user types the same question three different ways because the first answer wasn't quite right. Token visibility makes this category visible. Without it, you're just paying more.

What changes when you can see tokens per prompt.

1. Model selection becomes a habit, not a guess.

When the cost of each prompt shows up next to it in real time, people stop reaching for the most expensive model out of habit. You watch yourself try Sonnet first, see the answer is fine, and skip the Opus escalation 80% of the time. That single behavioral change cuts most teams' AI spend by 30–50% without losing output quality.

2. Long prompts get short.

AI cost scales with prompt length. The 8,000-token "here is everything about our codebase" preamble that prefixes every Cursor question? That's $0.024 per send on Sonnet, every single time. Twenty sends a day, twenty working days a month: $9.60 / month / developer just on the preamble, before they've asked anything. Token visibility makes that fat obvious. You see the number; you trim the prompt.

3. The retry tax becomes visible.

The single biggest hidden cost in AI usage is the retry loop. You ask a question, the answer isn't great, you rephrase, you ask again, you rephrase, you ask once more. Three tokens-charges for one answer. Once you can see the retry pattern in your data, you start writing better first prompts — which is the single highest-leverage skill anyone using AI can develop.

4. Budget anomalies stop being mysteries.

Last Wednesday's spend was 3× normal. Without per-prompt data, the conversation is: "did we use it more? did prices change? was there a bug?" — and you'll never know. With per-prompt data: "Marie ran a 4-hour Claude session generating SQL migrations against a 20k-line schema. That's the spike." Twenty seconds to root cause.

Privacy is the catch — and it has a clean answer.

The objection most teams raise the first time they hear "we'll track every prompt" is reasonable: those prompts contain our code, our customer data, our internal strategy. That's why the right model for AI usage tracking is token counting in the browser, never prompt text leaving the device.

A browser extension can count tokens on the page locally and emit only the totals — input count, output count, model, timestamp — to the analytics layer. The prompt content itself never travels. That's the pattern TokenEyez uses, and it's the pattern anything you adopt for this should use too. If a tracking tool wants the full prompt bodies, walk away. The token counts alone give you all the visibility; the prompt bodies give you legal exposure.

Treat tokens like you treat cloud spend.

The mental model that works: tokens are the new EC2 instance-hours. Free to read about, ruinous if you stop paying attention. Cloud teams learned, after a few hard quarters, that the cost dashboard isn't pessimism — it's instrumentation. The same shift is happening with AI now, just faster, because the bills grow faster.

The team that ships best with AI in 2026 won't be the one with the biggest model budget. It'll be the one that sees the model budget — per prompt, per developer, per project — and adjusts daily. Visibility first. Optimization second. Cuts last.

That's why we built TokenEyez. It's also why you should care, whether you use us or build your own.

For the broader context on why this matters now — and why Nvidia's CEO argued that engineers should consume tokens worth half their salary — read Tokens are the new productivity metric. If you'd rather just ask plain-language questions about your own spend from inside Claude or Cursor, the TokenEyez MCP server is the 90-second route.