API Error 500 at ~175K context — 1M window not active on Max plan (Opus 4.6)

Summary

During an active Claude Code session on a Max subscription, the API returned an HTTP 500 Internal Server Error at ~175K tokens of context. The 1M context window announced as GA for Max subscribers appears to not be active — context crashed near the legacy 200K boundary instead of growing to 1M.

Note: Claude Code auto-updated to v2.1.104 shortly before this session. The update may be related.

Environment

| Field | Value |
|-------|-------|
| Claude Code version | 2.1.104 (recent auto-update) |
| Model | claude-opus-4-6 |
| OS | Linux 6.17.0-20-generic (Ubuntu 24.04, x86_64) |
| Entrypoint | CLI |
| Plan | Max |

Affected Session

| Field | Value |
|-------|-------|
| Session ID | ca2021fd-6968-4fb2-b2f4-efbcea6b90f7 |

Timeline

| Time (UTC) | Event | Context (tokens) |
|------------|-------|-----------------|
| 14:13:22 | Session start | 27,723 |
| 14:29:57 | Normal growth | 74,296 |
| 14:45:50 | Normal growth | 100,480 |
| 15:15:41 | Normal growth | 122,648 |
| 15:26:43 | Normal growth | 147,949 |
| 15:31:42 | Last successful call | 174,857 |
| 15:31:48 | API Error 500 | 0 (synthetic) |

• Total API calls: 186 (185 successful, 1 fatal error)
• Total output tokens: 108,029
• Peak context: 174,857 tokens
• Service tier: standard (consistent across all 185 calls — no long-context tier observed)

Error Details

{
  "type": "error",
  "error": {
    "type": "api_error",
    "message": "Internal server error"
  },
  "request_id": "req_011Ca21cQ4aEV3A2LGyu4H9d"
}

Key Observations

1. Context never exceeded ~175K — crashed near the legacy 200K ceiling, well within 1M limit
2. Zero compaction events — context grew monotonically 28K → 175K, then hard crash (no graceful degradation)
3. service_tier: "standard" on all 185 calls — no indication of 1M context being provisioned
4. Post-error degradation — subsequent sessions behave as if running with reduced context capacity
5. Recent auto-update to v2.1.104 — happened shortly before this session, may have affected context window configuration

Expected vs Actual

| | Expected (Max + 1M GA) | Actual |
|--|----------------------|--------|
| Context ceiling | 1,000,000 tokens | Crashed at ~175K |
| Behavior at 175K | Continue growing | API Error 500 |
| Service tier | Long-context tier | standard only |

Context Growth Chart

175K ┤                                                          ●✖ ERROR 500
170K ┤                                                      ●●●●
165K ┤                                                   ●●●
160K ┤                                                ●●●
150K ┤                                          ●●●●●●
140K ┤                                      ●●●●
120K ┤                               ●●●●●●●
100K ┤                        ●●●●●●●
 75K ┤               ●●●●●●●●
 50K ┤      ●●●●●●●●
 28K ┤●●●●●●
     └────────────────────────────────────────────────────────────
     14:13                    14:45          15:15      15:31 UTC

Questions for Anthropic

1. Investigate request req_011Ca21cQ4aEV3A2LGyu4H9d — root cause of the 500?
2. Is 1M context active for this account? service_tier: "standard" suggests it may not be provisioned
3. Did the v2.1.104 auto-update change context window behavior?
4. Is there a silent fallback to 200K on error? How does the user recover full 1M?

---
Data extracted from Claude Code session JSONL logs on 2026-04-13