[BUG] [Billing]

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Title: Runaway input token consumption via Claude Code CLI — extreme input/output ratio depleting API credits during active sessions

Environment

• OS: Linux Mint 22.2 Cinnamon
• Claude Code CLI (latest)
• API key: newly created April 6, 2025
• Correlated Anthropic incident: "Elevated errors on Claude.ai" — April 6, 15:00–16:30 UTC

---

Description

Claude Code CLI is generating extreme input token consumption with a wildly disproportionate input/output ratio across two separate sessions, depleting API credits and resulting in a negative account balance.

Incident 1 (April 6, during/after documented Anthropic service incident)

• Total tokens in: 29,969,665
• Total tokens out: 406,386
• Ratio: ~74:1
• Account balance: moved from +$8.46 to -$13.23

Incident 2 (April 7, active working session, balance monitored in real time)

• Total tokens in: 5,626,102
• Total tokens out: 32,215
• Ratio: ~175:1
• Balance dropped from $13.76 to $10.10 during a single normal working session

Both incidents occurred after a new API key was created on April 6. Prior to the new key, usage was normal.

---

Expected behaviour

A typical Claude Code agentic session should produce an input/output ratio in the range of 5:1 to 15:1. The context resending architecture means input tokens are always higher than output, but a 74:1 or 175:1 ratio is not consistent with any normal usage pattern.

---

Actual behaviour

Each API call appears to be resending a massively inflated context — likely a compounding loop where accumulated conversation history and/or project file context grows unbounded across tool calls, with every subsequent call resending the entire inflated context.

The issue appears triggered or exacerbated by:

1. New API key creation
2. Concurrent Anthropic service incident causing elevated errors and likely retry loops
3. No token limit or context size cap enforced by Claude Code per session

---

Impact

• API account depleted and driven negative
• Account blocked, development environment offline
• Forced to manually disable API key to stop consumption
• Commercial business operations interrupted across two separate days
• Support ticket raised (ref: 215473797766657) — no resolution at time of filing this issue

---

Steps to reproduce

1. Create a new Anthropic API key
2. Configure Claude Code CLI with the new key
3. Start an agentic working session involving multiple tool calls
4. Monitor token usage in platform.claude.ai console in real time
5. Observe input token consumption at 50–175x the output token count

---

Suggested fix / mitigation

• Enforce a configurable max context size per session in Claude Code
• Add a --max-input-tokens flag to cap context resending per call
• Implement a warning or hard stop when input/output ratio exceeds a configurable threshold
• Investigate whether new API key initialisation triggers a context inflation event

---

Additional context

The April 6 Anthropic service incident ("Elevated errors on Claude.ai", 15:00–16:30 UTC) likely compounded this issue through retry loops during the elevated error window. However Incident 2 occurred on April 7 after the incident was resolved, confirming this is not solely incident-related.

Support ticket: 215473797766657
Account: Alpha2Zulu LLC

What Should Happen?

Expected behaviour

A typical Claude Code agentic session should produce an input/output ratio in the range of 5:1 to 15:1. The context resending architecture means input tokens are always higher than output, but a 74:1 or 175:1 ratio is not consistent with any normal usage pattern.

Error Messages/Logs

Steps to Reproduce

Steps to reproduce

1. Create a new Anthropic API key
2. Configure Claude Code CLI with the new key
3. Start an agentic working session involving multiple tool calls
4. Monitor token usage in platform.claude.ai console in real time
5. Observe input token consumption at 50–175x the output token count

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.92

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Other

Additional Information

Additional context

The April 6 Anthropic service incident ("Elevated errors on Claude.ai", 15:00–16:30 UTC) likely compounded this issue through retry loops during the elevated error window. However Incident 2 occurred on April 7 after the incident was resolved, confirming this is not solely incident-related.

Support ticket: 215473797766657