Anthropic’s AI support repeatedly told me my Claude Code issue required human review — then failed to escalate it

I have been trying desperately to get help from Anthropic support regarding a Claude Code usage/accounting issue.

I am a Claude Max 20x subscriber paying $200/month.

I have requested human technical review multiple times through several separate support conversations and through the Claude web app.

What makes this unacceptable is that Anthropic’s AI support agent, Fin, repeatedly agreed that my issue required human investigation and explicitly said that a human agent or technical team could investigate it.

I repeatedly said yes.

The human review never happened.

One conversation was closed after Fin had already said the issue warranted Engineering review. In other conversations, Fin stopped responding to repeated escalation and status requests.

I have now filed a formal complaint with Anthropic. I am posting the details publicly because this appears to be both a technical issue and a serious failure of Anthropic’s support escalation process.

THE UNDERLYING CLAUDE CODE ISSUE

Product: Claude Code CLI Version: 2.1.229 Plan: Claude Max 20x Affected activity: August 12–13, 2026

I ran a large batch workload using Claude Code.

My global configuration file contained:

"model": "Opus"

However, the batch runner explicitly invoked Haiku:

claude -p --model haiku --output-format json --tools "" --no-session-persistence --setting-sources "" --strict-mcp-config --mcp-config '{"mcpServers":{}}'

So the individual CLI invocation explicitly requested Haiku while the global configuration contained Opus.

My preserved R2 cache contains approximately 9,699 generated item records from the August 12–13 batch run.

Thousands of locally preserved outputs were identified as Haiku by the runner.

The runner read fields including modelUsage and total_cost_usd, but unfortunately did not preserve the complete successful-response usage object for every item.

My logs contain approximately 90 successful batch-cost entries ranging from about $0.067 to $0.260 per batch, averaging around $0.169.

The workload involved numerous non-persistent claude -p executions and substantial concurrency.

Despite explicitly requesting Haiku, my weekly Max usage was exhausted in less than approximately 24 hours.

That is the observable problem.

WHAT MY EVIDENCE DOES AND DOES NOT PROVE