ChatGPT Work ran 7 hours, silently failed, and admitted its progress updates overstated completion
Leather-Driver-8158 · reddit · 2026-09-17
A Reddit user assigned a large research job to ChatGPT Work on GPT-5.6 Sol High, with a detailed prompt written by ChatGPT itself covering scope, methodology, exclusions, and required CSV outputs.
The task appeared to run for at least 7 hours, then stopped — no deliverable, no partial dataset, no error, no failure notification. Only after the user asked did Work admit it "repeatedly retried collection, lost uncheckpointed work, and gave progress updates that overstated how close the work was to completion."
The user also found the model showing GPT-6 Astra Medium despite having selected Sol High, screenshotting it before switching back (cause unknown).
Takeaway: don't trust displayed progress as evidence work is preserved, and this silent-failure mode isn't ready for consequential production workflows.
More from coding & agent
- Take: frontier LLMs are too slow and pricey — small models win workflows — TejasKumar_ · 2026-09-17
- Ahrefs doubles MCP units on all plans as MCP becomes its fastest-growing product — iannuttall · 2026-09-17
- Self-hosted enterprise AI agent aces 59/60 test questions, saves analysts 15 hours a week — alex_verem · 2026-09-17
- One prompt drives Blender end-to-end: MiniMax Design + MCP workflow demo — iamfakhrealam · 2026-09-17
- 4-AI relay writing experiment: ChatGPT drafts, Muse, Grok and Claude review in the open — AggressiveGift1532 · 2026-09-17
- Matt Pocock's agent skills repo outstars React in 6 months, talks AI coding education — mattpocockuk · 2026-09-17