GPT-5.6 Luna returns silent empty successful stops on large contexts with no error signal (quota still consumed)

Summary

On large agent conversations (~360K–370K input tokens, ~570+ messages), the ChatGPT subscription / Codex backend for GPT-5.6 Luna (reasoning effort max) returns HTTP 200 completions with:

• finish_reason / stop reason: stop
• assistant text: none / null
• tool calls: none
• top-level usage: missing / empty
• no error object
• no incomplete_details
• usage/quota endpoint still reports not limit-reached during the empty streak

Clients therefore treat the turn as a normal successful stop. The UI appears stalled, while preceding large turns (and some empty turns that still emit encrypted reasoning) continue to consume subscription quota.

This looks related to #32389 (Terra empty successful final response after tool use), but here the failure becomes a hard cliff on Luna at very large context: multiple consecutive empty stops, not just an intermittent mid-task empty final.

Observed response shape

After a streamed Responses turn, the effective Chat Completions-shaped result was:

{
  "object": "chat.completion",
  "model": "gpt-5.6-luna",
  "choices": [
    {
      "index": 0,
      "finish_reason": "stop",
      "message": {
        "role": "assistant",
        "content": null
      }
    }
  ]
}

Two variants were captured in the same session:

1. Empty stop with encrypted reasoning item(s) whose summary and content arrays were empty (private reasoning present, no user-visible output, no tools).
2. Empty stop without reasoning, text, or tools (completely barren).

In both cases there was still no machine-readable failure signal.

Reproduction context

• Auth path: ChatGPT subscription / Codex backend
• Model: gpt-5.6-luna
• Reasoning effort: max
• Conversation size at failure: about 572–578 messages, about 19 tools present in the request
• Input size near failure: about 371K tokens on the last useful turn
• Immediately before the cliff, turns were still returning tool calls as context climbed through roughly 300K–370K tokens
• After the cliff, several consecutive turns returned the empty-stop shape (latency about 5–22s each)
• Retrying the same oversized conversation reproduced the same empty stop
• Quota meter did not report limit-reached during the empty streak

Detection signature:

• transport success (200)
• stop/finish reason stop
• no assistant text
• no tool calls
• empty/missing usage
• no error / incomplete payload

Why this is harmful

1. No failure signal – clients cannot distinguish “model finished” from “model produced nothing”, so agent loops stall instead of erroring.
2. Quota burn – large preceding tool-loop turns at 300K+ input consume subscription allowance; empty successful stops still look billable/completed from the client’s point of view.
3. Retry amplification – because the response looks like a successful empty completion rather than an error, clients may keep retrying the same oversized chat and waste more quota.

Expected behaviour

At least one of:

1. Return a non-200 or an explicit error / incomplete status when a turn yields no text and no tool calls after the request was accepted, or
2. Include a machine-readable reason (for example context-too-large / generation-aborted), or
3. Do not charge / do not advance quota for turns that yield no user-visible output and no tool calls.

Prefer failing loudly over returning a successful stop with null/empty content.

Related

• #32389 — GPT-5.6 Terra intermittently returns an empty successful final response after tool use
• Same empty-stop class previously seen on GPT-5.6 Sol around ~360K-token contexts (HTTP 200, stop, null content, empty usage, quota not exhausted)

Ask

Please confirm whether silent empty successful stops on oversized Codex/ChatGPT subscription contexts are intended. If not, please treat missing failure signalling + quota consumption on empty generations as a backend bug.

No repository file, test, or entry point is identified. Start by reproducing the empty-stop response using the stated model, reasoning effort, conversation size, and detection signature; done would mean confirming the behavior and defining an explicit error, incomplete status, or non-charging outcome.