Anthropic’s S-1 was written for investors. Engineering leaders should read it anyway.

Behind the growth numbers are the budget questions headed straight for your next planning meeting: What are we getting for this spend? Is the cheaper model actually cheaper? How much of our workflow depends on a vendor whose decisions we don’t control?

Those questions already have millions of dollars riding on them. Nearly 2,000 customers spent at least $1 million with Anthropic over the past year. About 250 spent at least $10 million. And existing customers are ramping up spending at a staggering pace.

Somebody had to sign off on all those bills. And when spending blows past the plan, someone also has to explain what work all those tokens actually got done.

Token Engineering is how you keep that gap from eating your budget.

The one number freaking out your CFO

Anthropic reported roughly 900% net revenue retention. Existing customers spent about nine times as much as they did in the same quarter a year earlier.

No wonder finance teams are asking what the hell is going on! How much of that extra spending was in the budget at the start of the year?

And with nine times the spend, what should the company have to show for it?

If you’re still counting on cheaper models to slow that spending down, read on. The filing says Opus 5.5 cut customer cost per task by about 40% compared with Opus 5, but it also says those lower costs drive greater usage.

It makes a lot of sense. Your teams find more things to automate. Agents run more often, on more tasks. Yes, you’re paying less for each task, but now your teams are using AI for so much more work that the total bill keeps growing.

That’s why you need to know what it costs to get a change merged, an incident resolved, or a migration finished. Cost per outcome tells you whether that bigger bill is buying better results or just more activity.

Quality belongs in the budget

Our Speed Trap report found that as AI adoption deepened inside engineering teams, restarts rose 66.7% and time in QA rose 300.6%. More work was reaching production, but teams were spending longer proving it worked and going back to fix it.

That work costs money, too!

Our report shows why the token bill alone can’t tell you whether you’re saving money. If a model cuts inference costs but leaves engineers doing more review, testing, and cleanup, you have to count those hours. Will switching models help, or give them even more to do?

The S-1 itself cautions that industry benchmarks may not reflect real-world performance. A model can look great on a leaderboard and still leave your team with hours of cleanup. So, build an evaluation set from your own work.

Then compare the models, and count the cleanup.

Their limits are your limits

Anthropic’s risk factors say compute costs could rise sharply, and its cloud partners could charge more or give it less capacity. Anthropic has also changed usage and rate limits before, and the filing notes that customers built on older models may resist migrating when those models are retired.

That’s a lot of your budget and delivery plan riding on decisions somebody else gets to make. If pricing changes, a model is retired, or a rate limit starts blocking work, can you move? How long would it take? What would you lose?

You should know those answers before you need them!

Take one important workflow and run it on a second model. Measure the difference in cost, quality, and the work required to switch. You may decide your current model is worth every dollar. Fine. But if switching would take three months and derail your roadmap, that dependency belongs in your planning.

You still have to explain the bill

Anthropic has a clear story for investors: make AI cheaper to use, give it more work to do, and grow customer spending.

Engineering leaders need an equally clear story for why that spending is worth it.

If your AI bill doubles next year, what outcomes should the company expect in return? Which work gets done sooner? What can your team take on that it couldn’t before? And how much review, rework, and operational cost comes with it?

Know what that money needs to deliver. Then go build something worth the bill!