OpenAI Silently Capped ChatGPT Plus Extended Thinking at 26 Minutes Around August 20
Around August 20, 2026, OpenAI changed how long ordinary ChatGPT Plus sessions can run extended thinking tasks. What was once an execution window that could stretch beyond 100 minutes is now consistently capped near 25-26 minutes — and the change was made without announcement.
The Evidence
The documentation is technical and specific. One Plus subscriber recorded a preserved ChatGPT Plus ordinary-Chat turn of 6,146 seconds (102 minutes 26 seconds) running GPT-5.6 Thinking/Extended before August 20. After the change, the same account produced runs clustering tightly between 1,541 and 1,564 seconds (approximately 25 minutes 40 seconds to 26 minutes 4 seconds).
A second independent specimen measured 1,564 seconds. A fresh reproduction came in at 1,550 seconds. The clustering is not noise — at least one user reported the 25-26 minute boundary appearing across ten or more separate attempts on multiple projects. Another user reported prior runs reaching 107 minutes before the same regression appeared.
A key technical detail from the investigation: the ~26-minute limit tracks reasoning execution time, not browser connection lifetime. In one specimen, the SSE connection stayed open for an additional 60 seconds after the reasoning work stopped at 26 minutes and completed with HTTP 200, confirming the boundary sits inside the model’s compute allocation, not in network infrastructure.
Two Execution Classes, One Account
The key finding is that the two execution classes still coexist on the same Plus account — they have just been routed to different product surfaces.
Before August 20, ordinary ChatGPT carried markers indicating it ran on what the investigation identifies as the “wfr_/SAServer” worker family — a durable execution infrastructure capable of long-horizon tasks. Current ordinary ChatGPT Plus sessions now run on a foreground class, identifiable by a UUID-based (non-Temporal) session architecture that terminates near 26 minutes.
ChatGPT Work, running concurrently on the same account, still receives the worker-family handoff. In one test, a Work session began while an ordinary-Chat SSE was still active — confirming the infrastructure exists but that ordinary Chat no longer qualifies for it.
The Tier Implications
The split is functionally a capability differentiation between ChatGPT Chat and ChatGPT Work. Extended thinking tasks that require sustained compute — large-scale document review, multi-hour research, complex code generation across a large codebase — now get cut off in ordinary Chat at the 26-minute mark with work unfinished. The same tasks dispatched through Work continue.
OpenAI has not classified the change as intentional or acknowledged it publicly. The forum thread documenting the regression (Community case #13832863) requested that if the cap is intentional product policy, OpenAI simply say so. If it is a regression, the case provides specific historical and current request IDs for investigation.
What This Means
For users running extended reasoning sessions in ordinary ChatGPT Plus, the practical limit dropped by roughly 75% at some point in the week of August 18-25. The change is not universal infrastructure loss — the long-run execution class is demonstrably still running for Work users on the same accounts.
If sustained, this functionally makes ChatGPT Work the required tier for any task that previously needed 30 minutes or more of extended thinking in ordinary Chat. The pricing and access difference between the two surfaces is substantial.