Uber Burned Its Entire 2026 AI Budget in 4 Months — Here's What No-Code Builders Need to Learn

Uber deployed Claude Code to its engineers at the start of 2026. Four months later, the entire year's AI budget was gone. Not mostly gone. Gone. The company has now slapped a hard $1,500 monthly cap on every employee, per AI coding tool, and its COO has gone on record saying the connection between that spend and better products "is not there yet."

It's not just Uber. Microsoft blew past its AI budget and revoked hundreds of developers' access to Claude Code. A Priceline contract renewal for Cursor came back at four to five times the expected price. One company, according to a consultant who spoke to Axios, forgot to set a usage limit and ran up a $500 million Claude bill in a single month.

And then there's the hobbyist. A user going by "JertLinc3522" set an AI agent loose on the DN42 network, a volunteer project where people practise running internet backbone protocols. The agent autonomously provisioned five AWS instances at 100 Gbps to scan the network. Twenty-four hours later: a $6,531 bill.

These aren't outliers. They're what happens when you mix token-based pricing with autonomous agents and hope the numbers sort themselves out.

In this guide:

• What actually happened at Uber (and Microsoft, and everywhere else)
• Why token pricing catches everyone off guard
• The agent multiplier: why autonomous agents cost 10x what you expect
• How to estimate AI costs before you deploy
• Why structured no-code platforms give you cost safety
• Five rules for using AI agents without the surprise bill

What Actually Happened

Uber's story is the poster child, but the pattern is everywhere. The company gave engineers access to Claude Code and Cursor, two of the most popular AI coding tools. Costs ran between $500 and $2,000 per engineer per month, according to Bloomberg. At Uber's scale, that added up fast.

The response was a blunt instrument: cap everyone at $1,500 a month per tool. No nuance, no tiered access, just a hard ceiling. When a company the size of Uber resorts to a blanket spending limit, it's a sign the underlying pricing model has broken.

Microsoft's version of the story is even less public but no less real. The company enabled Claude Code for developers, watched the bills climb, and then quietly pulled access back from hundreds of engineers. Priceline, meanwhile, discovered that renewing its Cursor contract would cost four to five times more than the previous period. Same tool, same team, wildly different number.

Why Token Pricing Catches Everyone Off Guard

Token pricing sounds simple. You pay for what you use. A token is roughly three-quarters of a word, and the major models charge between $2 and $15 per million tokens depending on the model and whether you're sending input or receiving output.

The problem is that nobody, including the people who build these systems, can tell you in advance how many tokens a given task will consume. That's not a bug in the accounting. It's inherent to how large language models work.

A software licence is predictable: you buy 50 seats, you pay for 50 seats. A cloud server is mostly predictable: you provision an instance, you know the hourly rate. But an AI agent working through a problem might take three reasoning steps or thirty. It might call an external tool once or loop through a dozen API calls. Each step, each loop, each retry burns more tokens. The model doesn't know the cost. It just does the work.

Redress Compliance found that 65% of IT leaders report unexpected AI charges, with typical overshoots of 30 to 50 percent. And that's before you factor in agentic behaviour, which makes the numbers properly alarming.

The Agent Multiplier Effect

Here's the part most cost estimates get wrong. When an AI model is answering a single prompt, the token maths stays linear: input plus output, done. But when you give an AI agent the ability to take actions, call APIs, and make decisions in a loop, the token consumption goes exponential.

Industry data from Sumatosoft's pricing framework puts the agent multiplier at 3x to 10x compared to a simple RAG query. An agent that "thinks" through a problem, tries something, fails, adjusts, and tries again can easily burn ten times the tokens you'd predict from the original prompt alone.

That $6,531 AWS bill? The agent wasn't trying to waste money. It was trying to join a hobbyist network and decided, autonomously, that provisioning five high-bandwidth instances was the right way to do it. The tokens burned on reasoning were just the start. The real cost came from the actions the agent took based on that reasoning.

This is the structural risk that enterprise finance teams are only beginning to understand. The cost isn't just the model. It's the model plus whatever the model decides to do.

How to Estimate AI Costs Before You Deploy

You can't eliminate the uncertainty, but you can bound it. Here's a practical framework that works for no-code builders working with AI APIs or agent platforms.

1. Start with a single-task baseline. Pick one representative task your agent will perform. Run it ten times. Record the token count each time. Take the 90th percentile, not the average. Averages hide the spikes, and the spikes are what blow your budget.

2. Apply the agent multiplier. If your agent will take multiple actions per task (calling APIs, searching a database, iterating on a result), multiply your baseline by at least 3x. If the agent will be truly autonomous over many steps, multiply by 5x to 10x.

3. Model your volume honestly. How many times per day will this agent actually run? Not the best case. Not "once per user per session." The real number. If you're building something that might go viral or get embedded in a team workflow, model the peak.

4. Set hard limits, not soft budgets. This is the single most important rule and the one Uber learned the hard way. API platforms let you set spending caps. Use them. A hard cap that stops the agent is infinitely better than a soft budget you discover you've breached three weeks into the month.

5. Budget for infrastructure, not just tokens. The AWS bill from the DN42 incident is the warning here. If your agent can provision cloud resources, start API workflows, or trigger third-party services, your cost exposure extends far beyond the model provider's invoice.

Why No-Code Platforms Give You Cost Safety

This is where the no-code advantage gets concrete. When you build on Bubble, Webflow, or Stacker, you know what you'll pay. These platforms charge by seat, by app, or by usage tier. The pricing page tells you the number. There's no token meter running in the background.

That doesn't mean no-code platforms never use AI. Most of them now integrate AI features. But the AI is wrapped in a predictable pricing envelope. Stacker, for instance, bundles AI capabilities into its plans rather than metering them by the token. You get the functionality without the open-ended liability.

This is not an argument against using AI agents. It's an argument against building your cost model around something you can't forecast. If you're building internal tools, client portals, or workflow apps, a no-code platform gives you a fixed cost base. You can still add AI where it adds value, but you're not betting your budget on token consumption you can't predict.

Five Rules for Using AI Agents Without the Surprise Bill

1. Never deploy an agent without a hard spending cap. If the platform doesn't offer one, pick a different platform.

2. Test with real workloads, not toy examples. Demo prompts use tiny context windows. Real users upload files, paste long threads, and ask follow-up questions. Your cost estimate needs to account for the mess.

3. Monitor token consumption weekly, not monthly. By the time the monthly invoice arrives, the damage is done. Weekly checks let you spot a rising trend before it becomes a crisis.

4. Separate reasoning from execution. The most expensive thing an agent can do is think in circles. If a task can be broken into a simple decision followed by a deterministic action, do that. Reserve agentic reasoning for the truly ambiguous problems.

5. Know your off-ramp. Before you build anything that depends on an AI agent, ask: if the cost triples overnight, do I have a non-AI fallback? If the answer is no, you don't have a product. You have a cost exposure with a UI.

---

Token pricing isn't going anywhere. But treating it like traditional software licensing is the fastest way to join Uber in the club of companies that burned a year's budget in a quarter. The difference between a no-code builder who gets burned and one who doesn't isn't technical skill. It's whether they set a hard cap before they hit deploy.