[BUG] Intermittent API Error 500 during runs (Feb 2026, Opus 4.6)

Bug Report

Description

Intermittent API 500 Internal Server Errors occurring during Claude Code sessions. The errors appear during normal usage runs and disrupt workflow.

Error Details

{
  "type": "error",
  "error": {
    "type": "api_error",
    "message": "Internal server error"
  },
  "request_id": "req_011CYPhsh4jvvwHEaPRuvLm5"
}

Environment

• Claude Code version: 2.1.45
• Platform: macOS (Darwin 25.3.0, Apple Silicon)
• Shell: zsh
• Provider: Anthropic API (direct, not Bedrock)
• Model: Opus 4.6 (claude-opus-4-6) — default, no ANTHROPIC_MODEL override
• Auth: OAuth
• Date of occurrence: February 22, 2026

Steps to Reproduce

1. Start a Claude Code session with default Opus 4.6 model via Anthropic API
2. Use normally (tool calls, multi-step tasks, subagent usage)
3. 500 errors appear intermittently during runs

What I've Ruled Out

• Not a model misconfiguration — using default Opus 4.6, no custom model env vars set
• Not Bedrock-related — running on direct Anthropic API (CLAUDE_CODE_USE_BEDROCK is not set)
• Not context-related — errors occur at various context sizes, including early in sessions
• Not network-related — other HTTPS traffic works fine; claude.ai web works simultaneously

Possibly Related

• There is an ongoing incident on status.claude.com from Feb 18, 2026 ("Intermittent errors in skills-related functionality") still in "Investigating" status as of Feb 22
• Multiple similar 500 error reports filed in February 2026: #22838, #22841, #23120, #16569
• Elevated errors on Sonnet 4.5 were reported and resolved on Feb 20

Expected Behavior

API requests should succeed or return actionable error messages (rate limit, context too large, etc.) — not opaque 500s.

Actual Behavior

Opaque 500 Internal Server Error with no actionable information for the user. Sessions are disrupted and the only workaround is to wait and retry.