Skip to content

claude-code-token-drain-bug.md

Latest commit

 

History

History
133 lines (85 loc) · 5.26 KB

claude-code-token-drain-bug.md

File metadata and controls

133 lines (85 loc) · 5.26 KB

Claude Code Is Burning Your Tokens 4x Faster Than It Should

Same projects. Same routines. Same time. But today it burns through your 5-hour limit 4x faster. Your quota drops while you're not even typing.

The Phantom Token Drain

From anthropics/claude-code#16856 (68 upvotes, 69 comments):

"I am working on the same projects for months, same routines, same time. But today it hits 5h limits like 4+ times faster! I am already on MAX and before today I would usually never hit even 50% weekly cap."

"Just waited for 5 hour reset, ran /usage and got 3% already used. Like what? Where did my 3% 5h limit just go?"

From #38239 (56 upvotes, 62 comments):

"Just had this happen to me. Used minimal prompts like I usually do, and within 30 minutes, I have already used up my 5-hour window. I'm on a max subscription."

From #6457 (30 upvotes, 119 comments):

"Something is definitely wrong with their usage limits calculation. I'm getting the same thing after literally only running a /compact command today."

Why Your Tokens Disappear

Three things are eating your quota silently:

1. The 1M Context Window Tax

Claude Code now supports 1M token context windows. This means every turn sends the entire conversation history to the API. A 30-minute session might have 200K tokens of context. Each prompt you send costs 200K+ input tokens even if your actual message is 10 words.

"The 1M context thing would make sense. Bigger window = more tokens per turn even if your actual prompts are the same size, since the full conversation gets sent every time."

2. Auto-Compaction Burns Tokens

When conversations get long, Claude Code runs /compact automatically. Compaction itself is an API call that processes the entire conversation. Users report hitting rate limits just from compaction running in the background.

3. Version-Specific Regressions

Multiple issues trace sudden token drain to specific Claude Code updates:

"Rolled back to 2.0.61 and my usage seems normal again."

Updates can change how aggressively context is sent, how often compaction runs, or how tools are invoked. Users have no visibility into these changes.

The core problem: you can't see what you're spending

Claude Max shows a usage percentage that:

  • Doesn't correspond to actual tokens used
  • Updates inconsistently
  • Sometimes shows low usage while you're rate limited
  • Has no breakdown of where tokens went

RunCode: Every Token Visible

RunCode shows you exactly what you're spending in real time:

╭──────────────────────────────────────────╮
│ Type a message...                        │
╰──────────────────────────────────────────╯
 anthropic/claude-sonnet-4.6  ·  $4.02 USDC  -$0.0234

That bottom line updates on every request. You see:

  • Your wallet balance ($4.02 USDC)
  • Session cost so far (-$0.0234)
  • The exact model being used

No hidden consumption. No phantom drain. If a compaction costs tokens, you see it. If context is growing, you see the cost increase per turn.

Smart token management built in

RunCode actively fights token waste:

  • Automatic context compaction when approaching limits -- but you see the cost
  • Tool result deduplication -- strips redundant output to save tokens
  • ANSI stripping -- removes terminal formatting from tool results
  • Per-request cost tracking -- every API call logged with exact token count

Switch to cheaper models when you don't need Claude

Most coding tasks don't need Opus. RunCode lets you use the right model for the job:

/model deepseek    # $0.28/$0.42 per 1M tokens -- 10x cheaper than Claude
/model flash       # $0.30/$2.50 per 1M tokens -- Google Gemini Flash
/model free        # $0.00 -- NVIDIA free tier
/model sonnet      # When you actually need Claude quality

A typical workflow: use DeepSeek or free models for routine tasks, switch to Claude for hard problems. This alone cuts costs 5-10x versus using Claude for everything.

The Comparison

Claude Code (Max) RunCode
Token visibility Usage % (unreliable) Exact token count + cost per request
Phantom drain Yes (background compaction, context) All costs visible in real time
Cost per session Unknown (hidden behind %) Shown in dollars
Compaction cost Hidden Visible
Model switching Not possible Instant: /model
Cheapest option $20/month (Pro, rate limited) $0 (free NVIDIA models)
Version regressions Can't detect Cost spike is immediately visible

Getting Started

npm install -g @blockrun/runcode
runcode

# Start cheap to see how tokens flow
/model deepseek

# Check your spend anytime
/stats

See exactly where every token goes. No surprises.

Links: