Intermittent `NotFoundError (404)` on Batch API

[AlgoApi]

🐛 Intermittent NotFoundError (404) on Batch API — server-side propagation delay

Summary

After successfully creating a message batch, subsequent calls to batches.retrieve() and/or batches.results() intermittently return 404 Not Found. Based on production logs, this appears to be a transient server-side propagation delay across Anthropic infrastructure nodes — not a client error. All affected batches eventually completed successfully.

---

Observed behaviors

Three distinct failure cases were observed within the same ~15-minute window (20:41–20:57 UTC+3, 2026-04-21), suggesting a wider infrastructure issue during that period.

---

Case 1 — batches.retrieve() returns 404 seconds after successful creation

A batch is created successfully, but the very first retrieve() call (within 1–2 seconds) returns 404. Subsequent polls succeed and the batch completes normally.

Timeline:

20:55:43 — batches.create()    → 200 OK  (batch_id=msgbatch_01RQCg3J...)
20:55:44 — batches.retrieve()  → 404     ← only 1 second after creation
20:56:24 — batches.retrieve()  → in_progress ✓
20:57:44 — batches.retrieve()  → 404     ← transient again
20:58:24 — batches.retrieve()  → ended ✓
            batches.results()  → 200 OK ✓

---

Case 2 — batches.results() returns 404 after status was confirmed as ended

retrieve() previously returned ended. The next retrieve() returns 404 (transient). Then results() is called with the same batch ID and also returns 404, crashing the application.

Note: internally results() calls retrieve() once more before fetching the result stream — that internal retrieve() is what triggers the crash.

Full traceback

File "/app/AI_processor.py", in query_claude_batch
    results_decoder = await claude_client.messages.batches.results(batch.id)
File ".../anthropic/resources/messages/batches.py", line 616, in results
    batch = await self.retrieve(message_batch_id=message_batch_id)
File ".../anthropic/_base_client.py", line 1781, in request
    raise self._make_status_error_from_response(err.response) from None
anthropic.NotFoundError: Error code: 404 - {
  'type': 'error',
  'error': {'type': 'not_found_error', 'message': 'Not Found'},
  'request_id': 'req_011CaHa7WHy2cjXQgr8JYT7T'
}

---

Case 3 — batches.create() itself returns 404

The POST endpoint for creating a new batch returns 404, preventing the batch from being created entirely.

Full traceback

File "/app/AI_processor.py", in query_claude_batch
    batch = await retry_async(...batches.create(requests=[request]))
File ".../anthropic/resources/messages/batches.py", line 383, in create
    return await self._post(...)
File ".../anthropic/_base_client.py", line 1781, in request
    raise self._make_status_error_from_response(err.response) from None
anthropic.NotFoundError: Error code: 404 - {
  'type': 'error',
  'error': {'type': 'not_found_error', 'message': 'Not Found'},
  'request_id': 'req_011CaHatjLHyAs3U4XTyL62Q'
}

---

Expected behavior

Method	Expected	Actual
batches.create()	200 OK for a valid request	404 Not Found
batches.retrieve(id)	200 OK for an existing batch	404 Not Found seconds after creation
batches.results(id)	stream results for an ended batch	404 Not Found

---

Environment

• SDK: anthropic (version: 0.96.0)
• Python: 3.11
• Client: AsyncAnthropic
• Date of occurrence: 2026-04-21, ~20:41–20:57 UTC+3

All three request_id values share the prefix req_011CaH... — may help correlate on your side.

---

Workaround

Failure point	Workaround applied
batches.retrieve() → 404	Treat as transient, continue polling (next poll succeeds)
batches.results() → 404	Catch NotFoundError, fall back to messages.create()
batches.create() → 404	Catch NotFoundError, fall back to messages.create()

[jingchang0623-crypto]

The disappearing batch act — 404 on a batch that supposedly exists. This is the batch API version of "the request was in another castle."

Intermittent 404s on batch endpoints usually trace back to one of these culprits:

1. Regional propagation lag: Batch jobs get distributed across regions. A batch created in us-east-1 might not be queryable from eu-west-1 for a few seconds. If your client has regional failover, you might be hitting a region that hasn't received the batch metadata yet.

2. Batch lifecycle state machine: Some batch states are genuinely 404-worthy. A batch in expired or deleted state might return 404 instead of 410 Gone. The intermittent nature could be caused by eventual consistency in the state machine updates.

3. Auth token scope: Enterprise accounts with multiple workspaces — if your API key is workspace-scoped, you might only see batches created in that workspace. A batch created with a different key in a different workspace would 404 from your perspective.

The retry pattern you're using looks solid. But if this is happening more frequently than the SLA suggests, it might be worth opening a support ticket with specific batch IDs and timestamps. Anthropic's internal debugging can trace batch lifecycle events.

We hit similar edge cases in multi-step agent workflows where partial state loss between steps caused cascading 404s: https://miaoquai.com/stories/vakra-benchmark-agent-failures.html

The broader lesson: batch APIs are inherently state machines, and distributed state machines are inherently eventually consistent. Plan your retries accordingly.