If you've been using Claude Code and found your usage limit disappearing faster than you can blink, you're not alone. It's the number one reported issue with Claude right now. Users are watching simple requests drain their entire daily allowance in a single session — sometimes in minutes.
And here's the important part: this one is not your fault. This isn't a clarity problem or a prompting problem. This is Claude Code doing what it was built to do — just doing a lot more of it than you realised.
Building the plane while flying it
Some portions of Claude feel like they are building the plane while they are flying it. Claude Code is the best example of this. It is genuinely extraordinary technology — powerful, capable, and unlike anything that existed a few years ago. But like any new technology, it has teething issues. And right now, one of those teething issues is hitting a lot of people hard.
That's not a criticism. That's the honest reality of being at the frontier of something new. The question is understanding what's happening so you can work with it rather than against it.
What is a tool call and why does it matter?
When you ask Claude Code to do something — fix a bug, write a function, review a file — Claude doesn't just think and respond. It takes actions. It reads files, runs code, checks outputs, writes changes. Each one of those actions is called a tool call.
Think of it like hiring a tradesperson to fix a leaky tap. A straightforward job, you'd think.
But the tradesperson arrives and checks under the sink — tool call 1. Turns off the water — tool call 2. Inspects the fitting — tool call 3. Runs to the van for a part — tool call 4. Checks the pipe behind the wall while they're in there — tool call 5. Tests the water pressure — tool call 6.
Three hours and sixty tool calls later, your tap is fixed. The job got done. But you're paying for every single one of those actions — and nobody told you that's how the billing worked.
That's Claude Code. Every action it takes costs tokens. And it takes a lot of actions.
Reports from early 2026 showed Claude Code making 68 or more tool calls in a single turn on what seemed like a straightforward request. Each call re-reads the conversation, adds to the context, and burns through your allowance. Simple tasks that feel like they should cost one message can cost fifty.
Then there's thinking mode
On top of tool calls, Claude Code has a "thinking" mode where it reasons through problems before responding. This makes its answers significantly better — but it also uses a significant amount of tokens just in the thinking process itself, before it's even started the actual work.
Imagine asking someone a question and they spend forty minutes writing private notes to themselves before they answer you. The answer is great. But you're paying for the forty minutes of note-taking too.
The combination of thinking mode plus aggressive tool calling is what's draining budgets so fast. Neither feature is a problem on its own. Together, on a complex task, they can empty a daily allowance in a single session.
What you can do about it right now
Break tasks into smaller pieces
Instead of asking Claude Code to "build the whole login system," ask it to "write just the database schema first and stop." Smaller scope means fewer tool calls per request.
Use /clear between tasks
Typing /clear in Claude Code wipes the session context. A fresh context means Claude isn't re-reading everything from the beginning on every tool call. Dramatically reduces token burn on long sessions.
Be explicit about what you don't want Claude to check
Claude Code will explore broadly unless you tell it not to. "Fix only the login function, do not touch any other files" stops it going on an investigation tour of your entire codebase.
Work at off-peak times
During US business hours Claude's servers are under maximum load. Some users report noticeably higher token consumption during peak hours. Evenings and weekends are consistently more efficient.
Use Claude.ai for non-code tasks
Claude Code is built for coding. If you're using it for writing, analysis, or general questions — stop. Claude.ai handles those tasks at a fraction of the token cost with the same quality output.
The bigger picture
It would be easy to read all of this and feel frustrated with Anthropic. Don't. These are genuinely hard engineering problems at the absolute frontier of what AI can do. The fact that Claude Code exists at all — that you can describe a problem in plain English and watch a machine write working code, fix bugs, and reason through complex systems — is remarkable.
Yes, some portions of Claude right now feel like building the plane while flying it. The tool call billing isn't fully transparent. Thinking mode tokens aren't clearly communicated. Some sessions cost far more than users expected and that's a real problem that needs fixing.
But look at what Anthropic has actually built. A tool that can write code, reason through problems, understand context, and work across industries in plain English. Any new technology has teething issues. That's not a flaw — that's the cost of being first.
What a plane Anthropic has built. We'd fly that plane any day.
If Claude Code's token costs are hitting your budget harder than expected, a session with us can help you set up smarter working habits around it — smaller tasks, better scoping, and a CLAUDE.md file that stops it going off-script. Most clients cut their token usage significantly in the first week.