Silent 429 on context-1m-2025-08-07 beta after model switch — no rate-limit headers, no retry-after

Summary

Switching from claude-sonnet-4-6 to claude-opus-4-7 with the context-1m-2025-08-07 beta (/model claude-opus-4-7[1m]) in a session with an existing Sonnet prompt cache causes the first request to trigger a persistent HTTP 429 that:

1. Returns no anthropic-ratelimit-* headers
2. Returns no retry-after header
3. Body is only {"type":"error","error":{"type":"rate_limit_error","message":"Error"}}
4. Persists exactly ~71 minutes (matches ephemeral_1h TTL), regardless of client behavior
5. Blocks Opus/Sonnet; Haiku keeps working

Environment

• Claude Code CLI 2.1.114
• Auth: OAuth (Claude Max)
• Date: 2026-04-19, 15:26–16:37 UTC

Reproduction

1. Session with claude-sonnet-4-6, ~100k tokens of prompt cache built up
2. /model claude-opus-4-7[1m]
3. Next prompt → 429

Evidence

Last successful Sonnet request (15:23:57 UTC):

{"model":"claude-sonnet-4-6","usage":{"cache_read_input_tokens":106561,"ephemeral_1h_input_tokens":258}}

Model switch 15:24:16 UTC → first 429 at 15:26:00 UTC:

HTTP 429, headers: {} (empty), body: {"type":"error","error":{"type":"rate_limit_error","message":"Error"},"request_id":"req_011CaDT7QKiJjeuzkvXd2JNG"}

First successful Opus request (15:27:26 UTC) reveals the trigger:

{"model":"claude-opus-4-7","usage":{"cache_creation_input_tokens":209698,"cache_read_input_tokens":0,"ephemeral_1h_input_tokens":209698}}

→ 209,698 tokens of fresh ephemeral_1h cache creation because the Sonnet cache cannot be read by Opus. 60 consecutive 429s over 71 min, all empty headers, all generic "Error". Recovery auto-releases exactly at T+71min (ephemeral_1h TTL).

Request IDs (first/last): req_011CaDT7QKiJjeuzkvXd2JNG / req_011CaDTWPsKMpq3gpyGcSS5o

Hypothesis

The context-1m-2025-08-07 beta has its own enforcement layer that omits standard rate-limit metadata and releases only on ephemeral_1h slot expiry, not on volume drop.

Impact

• Undebuggable (generic "Error")
• Breaks SDK retry logic (no retry-after)
• 70+ min lockout from one oversized request
• Looks like a partial outage to users (Haiku works, Opus/Sonnet don't)

Suggested fixes

API:

• Include anthropic-ratelimit-* + retry-after on beta-enforcer 429s
• Specific error message: "1h prompt cache quota exceeded, retry in N seconds"

Client (Claude Code):

• On /model switch mid-session, either warn about cache-miss cost or auto-downgrade first request's ephemeral_1h to ephemeral_5m