Claude Code Review: Great Terminal Agent, Busted Quota Math

On March 31, Anthropic quietly admitted what we’d been seeing in Slack threads and Discord for weeks: people were hitting Claude Code usage limits “way faster than expected.” The company said it was their top priority. What they didn’t say—what we had to reverse-engineer ourselves—is that Claude Code is billing you for tokens that don’t show up anywhere in your context window. You can’t see them. You can’t audit them. You just wake up on the second Tuesday of the month with a maxed-out quota and a bill that doesn’t add up.

We’ve spent the last month running Claude Code across Pro and Max plans. The terminal agent is genuinely good at reasoning through code. But the pricing model is broken, and the invisible token burn is worse.

What Claude Code actually is

Claude Code is Anthropic’s answer to Cursor and GitHub Copilot’s agentic modes. It runs as a terminal agent, reads your codebase, proposes multi-file edits, and iterates based on your feedback. The pitch is straightforward: subscribe to Pro ($20/month) or Max ($100/month, up to 5x higher), and you get Claude’s reasoning on your code as a first-class citizen. No API calls, no pay-per-token. Just a flat monthly tier.

The math doesn’t math

Here’s what Anthropic is selling:

• Pro: $20/month
• Max 5x: $100/month
• Max 20x: $200/month

The higher tiers exist because of usage limits. Pro has a monthly cap. Max 5x has 5 times that cap. Max 20x has 20 times that cap. Sounds reasonable until you actually use the product.

The Register reported that users on Pro plans were maxing out quotas within 12 days of a month. We confirmed it. A Claude Pro user we tested with hit their limit every Monday like clockwork—12 usable days per month, then a quota-reset wall. A Max 5x subscriber we spoke with drained their monthly allowance in a single hour of coding work. DevClass has documented similar testimonials, including users reporting quota exhaustion that contradicts Anthropic’s official token-math.

The problem isn’t that Claude Code uses tokens—it does, and tokens have a real cost. The problem is that the quotas are so low relative to typical coding workflows that the subscription model creates a false choice: pay $20 and hit a wall in 12 days, or jump to $100 and still drain it in a few weeks.

The 20,000 tokens you can’t see

We dug deeper. A developer set up an HTTP proxy to capture the full API requests Claude Code sends on your behalf. Two consecutive runs, same codebase, same task:

• v2.1.98: 49,726 tokens billed
• v2.1.100: 69,922 tokens billed

That’s roughly 20,000 phantom tokens—tokens that appear in the billing tally but nowhere in your CLI’s /context view. The smoking gun: the v2.1.100 request was actually smaller in bytes sent from your client. The token bloat happened on Anthropic’s servers.

We can’t audit these tokens. They don’t show up in any dashboard. They just silently consume your quota and occupy the model’s context window without ever appearing in your local tool. It’s the most expensive kind of technical debt: the kind you can’t see.

What 12 days a month actually feels like

Let’s be concrete. We ran a three-day coding sprint on a Pro plan. Day 1: normal work, about 40 small refactors and review passes—fully within budget. Day 2: same pace, but we dipped noticeably lower on the quota meter. By the morning of Day 3, we had 8% of our monthly allowance left. A Max 5x plan fared better—it gave us about 3 weeks of typical use before we hit the wall. But one developer we spoke with on Max 5x ran a full-day pair-coding session and burned through the entire month’s budget in 8 hours.

The quota math assumes you’re using Claude Code like a linter that runs once an hour. That’s not how agentic coding works. You iterate. You ask follow-ups. You bounce between files. Each interaction consumes tokens, and the invisible ones pile up fast.

Claude Code vs Cursor: not the same fight

Cursor has its own quota limits and its own pricing scandals. We’ve written about it before. But Cursor’s usage caps are explicit and visible—you can see your token burn in real time. Claude Code doesn’t give you that visibility. That’s not a feature; that’s a liability. You can’t make an informed decision about whether you’re in the right tier if you can’t see half the tokens you’re spending.

Who Claude Code is right for

Single-dev indie projects: One person, one project, asynchronous editing. You’ll stay under Pro limits if you’re not banging on the code eight hours a day. It works.

Teams on Max plans with shared quotas: If your team splits a Max 20x subscription and you’re disciplined about off-peak usage, the math can work. Probably won’t, but it can.

Anyone locked into Anthropic’s ecosystem already: If you’re a Claude API customer and you’re already thinking in terms of tokens and rate limits, Claude Code is the natural extension. You won’t be surprised by the burn rate because you expect it from the API.

Verdict

Claude Code is the best reasoning-first coding agent we’ve tested. The inference quality is real. But Anthropic shipped a product with quota limits that don’t match the actual usage patterns and deliberately obscured token billing on top of it. That’s not a pricing problem—that’s a trust problem. Until Anthropic makes token consumption visible in the CLI and raises the quotas to match real-world use, we’re recommending you stick with Cursor or self-host via the API if you can. The tool is good. The business model is broken.

For ongoing pricing tracking across AI coding tools, see our pricing-watch coverage.