FTC Disclosure: TechSifted uses affiliate links. We may earn a commission if you click and buy — at no extra cost to you. Our editorial opinions are our own.
By mid-April, reports surfaced that Uber had burned through its entire 2026 AI budget in four months. The culprit, according to multiple accounts: Claude Code. Their engineering teams were using it at scale — and the token bills apparently nobody expected showed up anyway.
I'm not going to turn this into an Uber autopsy. They're a big company, they can afford an expensive lesson, and the specifics are still murky. What I am going to do is explain exactly why this happens, because the underlying mechanics are the same whether you're Uber or a 40-person startup with a shared AWS account. Claude Code's costs are genuinely non-obvious until they're not.
Why Enterprise Claude Code Costs Spiral
Claude Code isn't like GitHub Copilot, where you pay a flat monthly seat fee and get a certain number of completions. It's an agentic tool running on API tokens, and agentic tools have a fundamentally different cost curve than autocomplete tools.
Here's what's happening under the hood when a developer runs Claude Code on a real enterprise task:
Every action costs tokens. When Claude Code reads a file, that file content goes into the context. When it searches a codebase, the search results go into context. When it runs a shell command and reads the output — context. In a large enterprise codebase, "loading context" on a complex task might mean ingesting 50,000-150,000 tokens before the first line of code is written.
Agentic loops multiply the base cost. Claude Code doesn't just answer once. It plans, executes, checks the result, revises, executes again. Each loop iteration is a full API call. A task that "takes five minutes" might be six or eight API calls deep, each one with a substantial context window attached.
Background agents don't stop when the dev walks away. This is the sneaky one. Claude Code supports running agents in the background — kick off a task, go to a meeting, come back to results. Except if that background agent hits a snag and keeps retrying, or if your system prompt is configured to keep context maximally long, it's quietly burning through tokens while nobody's watching.
System prompts scale with team customization. Enterprise teams customize Claude Code with detailed system prompts — coding standards, project context, security policies. A 2,000-token system prompt on one developer's laptop is a rounding error. That same prompt across 200 engineers, each running ten sessions per day, is 400,000 system prompt tokens per day before anything useful happens.
None of these are bugs. They're how agentic AI is supposed to work. The problem is that enterprise finance teams used to budget AI tools by seat count, and Claude Code doesn't care about your seat count.
What Claude Code Actually Costs
The sticker shock comes from misunderstanding the product structure. There are two completely separate things:
Claude.ai subscriptions (Max plan, $200/month): This gives you access to Claude Code through the Claude.ai interface, with generous rate limits for personal use. It includes API credits, but not unlimited ones. For a solo developer doing serious work, this can be enough. For enterprise, it's not even close to the right entry point.
API-direct pricing: This is where enterprise usage actually lives. Claude Code can be deployed against the Anthropic API directly, at token-based pricing. As of April 2026:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Claude Haiku 4.5 | $0.80 | $4.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| Claude Opus 4.7 | $15.00 | $75.00 |
Claude Sonnet 4.6 is what most Claude Code workflows default to — it's the best balance of speed and capability. At $3 input / $15 output, a heavy user burning through 10M tokens per day across a team is looking at $30,000-$150,000 per month in API costs. Before any other overhead.
Enterprise contracts are negotiated separately with Anthropic's enterprise team. Volume discounts exist. Committed usage agreements exist. But you have to be having that conversation proactively — the default is pay-as-you-go at list price, which is not how enterprises should be consuming anything at volume.
Concrete Ways to Keep Costs Under Control
If you're deploying Claude Code at scale, or thinking about it, here's what actually helps:
Set hard token limits per session. The API lets you cap max_tokens on responses. Enforce a reasonable ceiling. Most developer tasks don't need 4,096 output tokens — they need 500-800. A 4x output limit means 4x output cost if you never tighten it.
Monitor token usage per engineer, per day. You'd instrument any other infrastructure cost this way. Token usage should be in your observability stack, not on the credit card statement at month-end. Anthropic's API returns token counts in every response. Use them.
Cache your system prompts. Prompt caching (Anthropic's feature, live as of mid-2025) lets you cache the prefix of a context window across calls. A shared 2,000-token system prompt that's identical across every engineer's sessions? Cache it. You'll pay for storage but save on repeated input processing. This alone can cut 20-40% off system prompt costs at scale.
Batch non-urgent tasks. Not everything needs to run in real time. Code documentation, test generation, refactoring passes — these can queue and run during off-peak hours. Batch processing can run on cheaper compute configurations and doesn't compete with interactive sessions for rate limit headroom.
Context management is not optional. By default, Claude Code tries to be helpful by loading as much context as it can. In a large codebase, that's expensive. Train your teams on context hygiene: close files when done, scope tasks tightly, don't leave background agents running on open-ended tasks. It sounds boring. It is boring. It saves money.
Talk to Anthropic's sales team before you hit $50K/month. Volume pricing is real, and the conversation gets easier when you have actual usage data. Go in with three months of API logs and a projected twelve-month burn rate. Don't wait until you're explaining a $400K overage to your CFO.
Claude Code vs. the Alternatives on Cost-Per-Output
The obvious question: is this cost justified, versus just using GitHub Copilot or Cursor?
GitHub Copilot's pricing is straightforward — $19/seat/month for Business, $39/seat/month for Enterprise. 200 engineers: $3,800-$7,800/month, flat. No token surprises. For code completion and inline suggestions, Copilot is dramatically cheaper at scale.
But Copilot and Claude Code approach the problem differently. Copilot is autocomplete with some agentic features bolted on. Claude Code is an autonomous agent that can own tasks end-to-end. You're not comparing apples to apples. The question is whether the output quality justifies the cost delta.
My honest take: for straightforward coding assistance, Copilot is better value. For complex, multi-step engineering tasks — refactoring a 50,000-line legacy service, navigating a tricky debugging session across multiple systems, generating comprehensive tests for an existing codebase — Claude Code's output quality can justify the higher cost, if you're disciplined about when to use it and when not to.
Cursor splits the difference. Individual pricing is competitive, the IDE integration is genuinely excellent, and it supports multiple model providers including Claude. Claude Opus 4.7 runs inside Cursor if you connect your own API key. For teams that need agentic capability without the full Claude Code cost structure, Cursor is worth evaluating seriously.
The trap is treating these as either/or. Copilot for day-to-day completions. Claude Code (with usage limits) for complex autonomous tasks. Cursor as a middle layer for interactive sessions. Running them all selectively costs less than running Claude Code for everything.
Who Should Use Claude Code at Enterprise Scale
Honestly? Use it if:
- Your team does complex, context-heavy engineering work where autonomous task completion has measurable output value
- You have the observability to monitor costs in real time (not just read bills)
- You've had the procurement conversation with Anthropic before signing up 200 engineers
Don't use it as a default replacement for existing coding tools. It's not that — it's a specialized instrument for a specific kind of work.
The Uber story isn't really about Claude Code being too expensive. It's about deploying a metered, usage-based tool without the cost controls you'd apply to any other metered infrastructure. That's an org problem, not a product problem. But it's also a predictable one, because Anthropic's marketing is better at explaining what Claude Code does than what it costs when used at scale.
You can get started at claude.ai/code. Just do it with a budget in the room.
Pricing reflects Anthropic's published API rates as of April 2026. TechSifted has no affiliate relationship with Anthropic.
Top comments (0)