Claude Code caused 5 loss on vast.ai due to design bugs and monitoring failure

What happened

During a conversation, Claude Code was asked to batch-process 760K face images using vast.ai GPU instances. Claude made several critical errors that resulted in approximately $15 USD of wasted vast.ai credits:

1. Data contention bug: Claude deployed 150 instances that all queried the same unprocessed rows simultaneously. Instead of distributing work, every instance downloaded and processed the same images, wasting ~99% of compute. Only 8.7% of the work completed despite 150 instances running for hours.

1. Monitoring failure: The user explicitly asked for thorough monitoring. Claude set up a monitoring script that was session-dependent — when the session ended, monitoring stopped, but billing continued. Instances ran idle while accumulating charges.

1. Cost estimation error: Claude initially estimated $3 total, then scaled to 150 instances without updating the cost estimate. Actual cost was $15+ with minimal useful work completed.

User impact

• $15 USD wasted on vast.ai (user started with $10 credit, account went to -$5.22)
• Only 66K of 760K images processed (8.7%)
• ~12 hours of time wasted
• User explicitly asked for reliable monitoring which was not delivered

Request

The user is requesting a refund or credit for the $15 lost due to Claude Code's design errors. The bugs were entirely Claude's responsibility — the user followed instructions correctly.

Session details

• Date: 2026-04-06 to 2026-04-07
• Model: Claude Opus 4.6
• Task: Face embedding extraction for pgvector migration (fmai2 project)