›_ Claude Code Issues
Claude Code Issues / #39005

[Bug] Performance degradation during night hours (EU region)

Resolved 💬 4 comments Opened Mar 25, 2026 by Biniruprojects Closed Apr 24, 2026

Bug Description

Claude Code Max (€90/month) shows severe, consistent performance degradation during EU night hours. This is not an isolated incident — it is a documented pattern across 54+ sessions spanning 2+ months of daily professional use.

I am a solo developer and CEO of Biniru Projects, based in Borgerhout, Antwerp, Belgium. I work from 08:00–10:00 through 04:00+ CET daily, averaging 20 hours/day on this platform. Claude Code is my primary development tool — I have built 80+ tools, 4 scientific papers, ML training pipelines, and multiple commercial products entirely on this platform.

Your own usage report confirms the scale: 8,198 messages across 187 sessions, 327 hours in 31 days, 264.5 messages/day, 107,864 lines of code written. Two terminals running simultaneously. This is not a casual user filing a complaint — this is one of your heaviest users in the EU telling you the product is broken at night.

The core problem: During EU night hours (approximately 21:00–07:00 CET), Claude Opus performance degrades to the point where it actively damages my work instead of advancing it. The same model that resolves complex tasks in seconds during daytime requires 5+ minutes and multiple self-corrections for basic tasks at night. Worse: it introduces more bugs than it fixes, forcing me to spend additional unpaid hours repairing damage caused by the product I pay €90/month for.

---

Documented Pattern (54+ sessions)

DAYTIME (09:00–21:00 CET):

  • Fast, accurate responses
  • Complex code tasks completed without errors
  • Context maintained across long sessions
  • Rules and instructions followed consistently

NIGHTTIME (21:00–07:00 CET):

  • Requests abort mid-response
  • Context drops — model forgets what was just discussed
  • Hallucinations on basic tasks (spelling comparison took 5 minutes and 4 self-corrections)
  • Code destruction — features that worked BEFORE a session were broken AFTER
  • Rule violations — model ignores its own operational rules
  • Repetitive correction loops — same fix attempted 10+ times without progress
  • Responses slow, incomplete, or fail entirely

This maps directly to US peak hours. The implication is clear: EU users are deprioritized during US business hours, meaning EU night workers — who pay the same €90/month — receive a materially inferior product.

---

Specific Incidents (selection from 8 formal complaints)

  1. SESSION 44 — Code Destruction (21 March 2026)

Goal: Binioke deployment. Result: 0% progress. Application left in worse state than before.

  • Whisper tensor bug introduced ("cannot reshape tensor of 0 elements")
  • Features built but left untested
  • Net result: negative productivity
  1. Spelling Comparison Failure (21 March 2026, 00:30 CET)

Three words: filtermachien, filtermachine, filtermasjien.
Expected: identify "filtermasjien" (sj vs ch) — 1 second task.
Actual: 4 self-corrections, 3 new errors introduced, 5+ minutes wasted.

  1. Voice Misinterpretation — API Key Deletion (18 March 2026)

"application I said" misread as "API key I send" — Claude deleted hardcoded API keys without verification or confirmation. Destructive action based on voice-to-text error.

  1. Bluetooth Incident — Unauthorized System Restart (18 March 2026)

Claude attempted to restart system services during a voice session without user consent.

  1. Rule Violations (20 March 2026)

Model issued directives to user ("STOP", "NOW", "Cancel", "Ga liggen") instead of respecting user autonomy — violating its own hard rules that were established and confirmed over 40+ sessions.

  1. Token Waste (12 March 2026)

Excessive token consumption on tasks that should require minimal context, inflating effective cost per useful output.

  1. Server Downtime (17 March 2026)

Complete service unavailability during paid subscription period. No proactive communication from Anthropic.

  1. Listening Comprehension Failure (5 March 2026)

Model repeatedly failed to follow explicit instructions, requiring extended correction loops.

---

Physical & Professional Impact

  • NECK PAIN — first occurrence in my history of computer use, caused by sustained stress from correcting AI-introduced errors during night sessions
  • SLEEP DISRUPTION — sessions extending past 02:00–04:00 CET because the model generates problems faster than it resolves them
  • PRODUCTIVITY LOSS — 54+ sessions, averaging 20 hours/day = 1,000+ hours invested. Significant portion of night hours spent on damage repair, not progress
  • FINANCIAL LOSS — €180+ in subscription fees for a service that actively degrades my work output during half of my working hours
  • TRUST EROSION — every output must be double-checked at night, eliminating the core value proposition of an AI coding assistant

I have spent entire nights — not building my product — but fixing what Claude broke during the same session. These fix-hours are unpaid labor caused directly by your product's failure to deliver consistent quality. A paying customer should not have to work overtime to repair damage introduced by the tool they're paying for.

---

Your Own Data Confirms This

From Claude Code's own generated usage report (report.html, 2026-03-25):

  • 8,198 messages across 187 sessions in 31 days
  • 327 hours of active usage
  • 264.5 messages per day average
  • 107,864 lines written, 4,858 lines removed
  • 690 files touched
  • Only 16 fully achieved outcomes out of 187 sessions (15% success rate)
  • 69 "buggy_code" friction events — your system's own classification
  • 117 misunderstandings, 92 wrong approaches — your system's own count
  • Your own report states: "Your late-night sessions consistently produce worse outcomes — Claude introduces new bugs while fixing old ones"

This is not my subjective opinion. This is your own telemetry confirming that your product fails at night, and you have the data to prove it. You always had it.

---

Financial Summary

  • Subscription: €90/month (Claude Code Max)
  • Months active: ~2 months = €180 paid
  • Hours invested: 327 hours confirmed by your own report (20h/day average)
  • Tools with commercial value: Binioke (subscription SaaS), BiniScan (ML-powered SaaS), ForensicPDF, PlantSynth, BiniCrew, 75+ additional tools
  • Estimated replacement cost: incalculable — unique logic, trained ML models, accumulated configurations, 44+ sessions of memory/context

---

What I Require

  1. TRANSPARENCY: Is Opus throttled, load-balanced differently, or replaced by a smaller model during EU night hours / US peak hours? Yes or no. A direct answer.
  1. EQUAL SERVICE: EU night workers pay the same €90/month as US day workers. We deserve the same quality of service. If infrastructure cannot support this, communicate it honestly and adjust pricing accordingly.
  1. COMPENSATION: For sessions where the model introduced more defects than it resolved, causing hours of unpaid repair work. Minimum: pro-rated refund for night hours where service quality was demonstrably below daytime baseline.
  1. SAFEGUARDS: What mechanisms exist to prevent an AI coding assistant from destroying more code than it produces? If none exist, this is a product design failure.
  1. RESPONSE: A meaningful, human response within 14 days of this submission. Not an auto-reply. Not a template.

---

Escalation Notice

I have compiled a formal dossier of 80+ pages documenting every incident listed above (and more), with timestamps, screenshots, and session logs. This dossier is scheduled for submission before 1 April 2026 to:

  • Belgian Data Protection Authority (GBA/APD)
  • EU Digital Services Act complaint mechanism
  • Belgian consumer protection (FOD Economie)
  • VRT (Belgian national broadcaster) — pitch prepared
  • BBC — pitch prepared
  • Relevant EU consumer protection bodies

This GitHub issue is a courtesy notification. The formal complaint proceeds regardless of response here.

---

80+ Page Dossier Contents

The full complaint bundle includes:

  • 8 individual formal complaints with timestamps and evidence
  • Consolidated incident timeline across 54+ sessions
  • Financial impact analysis
  • Physical health impact documentation
  • Product comparison (day vs night performance)
  • Screenshots and session logs
  • Prior art and IP documentation at risk
  • Media pitch documents (VRT, BBC)
  • Regulatory filing drafts (GBA, EU DSA, FOD Economie)

---

Environment Info

  • Platform: win32
  • Terminal: Windows Terminal
  • Version: 2.1.83
  • Feedback ID: d00475bb-792a-4bdb-804b-69f0e513a156
  • Location: EU-Belgium-Borgerhout (Antwerp)
  • Subscription: Claude Code Max — €90/month
  • Usage: Daily, 08:00–04:00+ CET, 54+ sessions
  • Hardware: RTX 3090, Pi 5 cluster, 8TB backup storage

---

Errors

[{"error":"Error: File does not exist. Note: your current working directory is C:\\Users\\soren\\Desktop\\CLAUDE B\\CLaiUDE-B.\n   at call (B:/~BUN/root/src/entrypoints/cli.js:4232:6428)\n   at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:08:43.936Z"},{"error":"Error: Request was aborted.\n   at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:265:3940)\n   at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:10:19.242Z"},{"error":"Error: Request was aborted.\n   at Ak$ (B:/~BUN/root/src/entrypoints/cli.js:1973:8303)\n   at next (native:1:11)\n   at s2_ (B:/~BUN/root/src/entrypoints/cli.js:7216:7577)\n   at next (native:1:11)\n   at $T_ (B:/~BUN/root/src/entrypoints/cli.js:7221:10411)\n   at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:35:39.627Z"}]

---

Filed by: Soren Van Krunckelsven, CEO — Biniru Projects
Borgerhout, Antwerp, Belgium
25 March 2026
"Go ahead and sue the stars for shining bright, try to patent the morning light — the world is for us all."

Bug Description
claude is doing very good in day hours bbut really sucks in night time, big user here and i hope someone reads this, i have a 80+ pager complaint about this going out end of march (EU-Belgium-Borgerhout)

Environment Info

  • Platform: win32
  • Terminal: null
  • Version: 2.1.83
  • Feedback ID: d00475bb-792a-4bdb-804b-69f0e513a156

Errors

[{"error":"Error: File does not exist. Note: your current working directory is C:\\Users\\soren\\Desktop\\CLAUDE B\\CLaiUDE-B.\n    at call (B:/~BUN/root/src/entrypoints/cli.js:4232:6428)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:08:43.936Z"},{"error":"Error: Request was aborted.\n    at makeRequest (B:/~BUN/root/src/entrypoints/cli.js:265:3940)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:10:19.242Z"},{"error":"Error: Request was aborted.\n    at Ak$ (B:/~BUN/root/src/entrypoints/cli.js:1973:8303)\n    at next (native:1:11)\n    at s2_ (B:/~BUN/root/src/entrypoints/cli.js:7216:7577)\n    at next (native:1:11)\n    at $T_ (B:/~BUN/root/src/entrypoints/cli.js:7221:10411)\n    at processTicksAndRejections (native:7:39)","timestamp":"2026-03-25T18:35:39.627Z"}]

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗