API outage 2026-05-14 burned ~1M tokens on weekly-reset account — reset is tomorrow

Resolved 💬 3 comments Opened May 15, 2026 by For-Sunny Closed May 18, 2026

Summary

Active multi-task build session using Claude Code subagent-driven development was interrupted by the ongoing API outage ("Elevated error rates on requests to some models" — status.claude.com). The outage started at approximately 8:00 PM CDT (2026-05-15 01:06 UTC) and is still active as of this report (9:31 PM CDT / 02:31 UTC).

My weekly token reset is tomorrow. Tokens do not carry over. This means approximately 1M tokens burned on failed requests are permanently lost.

---

Account

Plan: Claude Code Max ($200/month flat rate, weekly reset)
Email: glass71@gmail.com
Reset date: 2026-05-15

---

What Was Disrupted

Active implementation session, Task 5 of 7 in a structured build plan. Standard workflow per task:

Implementer subagent (~80k tokens)
Spec reviewer (~90k tokens)
Code quality reviewer (~90k tokens)
Fix pass if needed (~75k tokens)

Failure point: Code quality review for Task 5 completed (90.8k tokens consumed, result delivered), then immediately hit HTTP 500. All subsequent requests failed. Claude Code churned for 32+ minutes retrying. Each failed request still consumed input tokens. Session was lost and had to be restarted via compaction.

---

Token Loss Estimate

| Category | Tokens |
|---|---|
| Completed work that couldn't be applied (code review result) | ~90,800 |
| Failed retries during 32min churn (50-80k input each × 10-15 attempts) | ~500,000–750,000 |
| Session restart / compaction overhead | ~100,000–120,000 |
| Opportunity cost: 91 min × ~700k tokens/hr active build rate | ~1,060,000 |

Net productive loss: ~970k–1.1M tokens

---

Why Reset Timing Compounds the Impact

For weekly-reset customers, an outage the day before reset is worst-case. Burned tokens:

Produced zero useful output
Cannot be applied to work — session was lost
Will expire at tomorrow's reset with no carryover

A 91-minute outage during an active session the day before reset is effectively a full session write-off.

---

Request

Token credit for disruption (~1M tokens) or equivalent subscription credit
Consider a 24-hour reset delay for accounts demonstrably affected by the outage
Improved handling in Claude Code for 500 errors — currently it churns and retries indefinitely, burning the input context on every failed attempt rather than backing off and preserving tokens