GPT-5.6 High Reverts to 26-Minute Cap on Plus as OpenAI Infrastructure Tiers Diverge
A thread in the OpenAI Developer Community documents a concrete GPT-5.6 High regression on Plus accounts: tasks that previously ran past 100 minutes are now terminating at approximately 25-26 minutes, consistent with the historical foreground execution timeout.
One user documented a thinking_effort: extended (High) run that built a 4,380-row dataset and audited summaries before the SSE connection — opened HTTP 200 — terminated with “Error in message stream.” Turn analytics logged 1,552 seconds (25m52s). The same account had previously run a comparable task for 1h42m26s, nearly four times as long.
The Execution Tier Divide
The cause identified in the thread: OpenAI runs two distinct execution classes, and Plus GPT-5.6 High is landing in the wrong one.
Foreground execution handles requests in a non-persistent process. The historical cap: ~25-26 minutes. Network-level inspection shows no wfr_* worker markers, no SAServer indicators, no stream_handoff.
Temporal / durable workers — identifiable by stream_handoff and conversation-turn-* markers in HAR captures — maintain job state across network interruptions and run substantially longer. This is the path Work and Codex accounts show for GPT-6 Pro and GPT-6 Astra tasks.
The specific regression: GPT-6 Pro in Work gets durable worker admission even in Temporary Chat mode. GPT-5.6 High on Plus does not show those markers, and falls to the foreground envelope.
Not a Hard Universal Limit
The community data is not uniform. Separate users report 9-minute, 16-minute, and 32-minute High runs. A successful resume on Plus extended one turn to ~47 minutes. One Plus user observed stream_handoff for GPT-5.6 High — so durable worker admission is not blocked at the account level for Plus, only uncommon.
The consistent pattern: Plus accounts with GPT-5.6 High land in foreground execution significantly more often than Work or Codex accounts running GPT-6 Pro, and foreground execution degrades reliably near the 25-26 minute mark.
The Practical Gap
The model remained productive throughout the documented failures — building correct rows, auditing summaries — until the connection terminated. The issue is not reasoning quality or context management. It is job durability.
The route to durable execution is Work or Codex, both of which cost more. Plus is already metered access to GPT-6 Astra in Work and Codex contexts. Community users argue that extending credit-based access to long-run execution in Chat would be the natural extension.
OpenAI has not acknowledged the September regression. The same execution ceiling was previously documented around August 20, when a silent routing change cut extended thinking runs on Plus. Whether this is the same mechanism reasserting itself or a separate degradation is unclear from the network traces available. What is clear: the gap between what Plus and Work can sustain for agentic tasks is widening, and it is determined by infrastructure routing rather than the model running behind it.