ChatGPT Work Went Down for 5+ Hours on August 31

ChatGPT Work — OpenAI's agentic task mode for paid subscribers — spent most of Monday afternoon, August 31, 2026 in a partial outage. OpenAI's status page logs the incident from 11:04 AM to 4:28 PM ET (~5 hours 24 minutes), with Plus users losing Work mode entirely during the worst stretch.

If you routed production workflows through ChatGPT Work that day, this is your reminder that hosted agent modes need fallbacks — same lesson as prior Work-mode incidents and the reliability gap explainx.ai tracks in why agents haven't gone mainstream.

TL;DR — timeline

OpenAI classified the event as Partial outage · ChatGPT Work seeing elevated errors and latency. After resolution, OpenAI published an incident write-up on the same status page — worth reading if you need root-cause detail beyond the live timeline.

What ChatGPT Work is — and why this outage type matters

ChatGPT Work is OpenAI's long-running agentic mode for paid subscribers — distinct from a normal chat turn. Work spins up threads that orchestrate subtasks, use tools, and continue across minutes or hours (mobile thread orchestration shipped in July). When Work mode fails, users lose:

• In-flight tasks — research loops, document drafts, multi-step automations
• Tool continuity — connectors to email, files, or (as of September) enterprise data warehouses
• The "set it and forget it" workflow Altman has described as a new kind of computer

Consumer chat can often fall back to a shorter answer. Work mode failure is closer to a CI pipeline stuck red — the job isn't done, and there's no partial artifact you can ship.

That tier split matters for internal runbooks: your Plus power users may have seen total Work absence while Team/Enterprise users saw degraded but partially functional behavior — verify against your actual subscription mix before post-morteming "everyone was down."

What users experienced

Per OpenAI status updates and third-party monitoring:

• Could not start new ChatGPT Work tasks
• Could not continue in-progress Work tasks
• Plus plan especially affected — Work mode listed as unavailable
• Elevated latency even when partial function returned

OpenAI's @ChatGPT account on X posted mid-incident: "ChatGPT Work isn't working right now. Sorry about that, we're working on a fix!"

Downdetector peaked at 1,100+ reports for OpenAI products during the window — smaller than Microsoft's same-day Outlook spike (~4,600) but enough to confirm user-visible failure, not a silent backend blip.

Status page granularity — six updates in one afternoon

OpenAI's incident 01M1C5M4K0WC8PPT0Z175RJA1E followed a familiar pattern:

1. Latency first — elevated latency identified before errors were named publicly
2. Error rate spike — "elevated errors for impacted services"
3. Plus-specific callout — Work mode unavailable for Plus tier
4. Task semantics — explicit language: users cannot start or continue Work tasks (not just slow responses)
5. Mitigation deployed — 4:01 PM ET monitoring phase
6. Recovery declared — 4:28 PM ET all impacted services recovered

The gap between mitigation (4:01 PM) and resolved (4:28 PM) is 27 minutes of "monitoring" — typical when a fix is deployed but engineers wait for error-rate graphs to flatten. Teams building on Work should assume ~30-minute uncertainty tails even after OpenAI posts green.

Same day, separate Microsoft Exchange issues

CNBC reported Microsoft Exchange Online degradation the same afternoon — "users may experience degraded functionality with various Exchange Online functionalities." Microsoft tracked it through its 365 status account.

No evidence links the two outages. But the coincidence matters operationally: teams running Grok Outlook connectors (full read-write access) plus ChatGPT Work agent tasks hit both email and agent surfaces degraded within hours — a realistic preview of multi-vendor agent stack fragility.

Monday afternoon stack failure — a tabletop scenario

Imagine a revenue ops team on August 31:

The lesson isn't "don't use OpenAI or Microsoft." It's never chain two hosted agent surfaces without a local or alternate-vendor escape hatch for the same critical path.

Pattern: Work-mode incidents are recurring in 2026

August 31 was not isolated. OpenAI's status history shows Work-specific degradation multiple times:

Hosted agents are young infrastructure. OpenAI's own classification — partial outage for a five-hour Work blackout — signals that Work mode is not tier-1 SLA territory the way core chat completion might be. Plan internal SLAs accordingly: if your workflow cannot tolerate five hours of unavailability, it cannot rely on ChatGPT Work alone.

What builders should do differently

Compare to August 27, 2026 — OpenAI logged a separate incident where Workspace Agents and ChatGPT Work on web/mobile saw increased error rates for 4h 14m. Work-mode reliability is a recurring theme, not a one-off.

Production runbook — when ChatGPT Work goes red

For loop-engineering teams: treat Work mode like an external API with best-effort SLA, not like local while true you control.

Enterprise and security angle

Work outages aren't only UX problems. If your org connects Work to Snowflake, BigQuery, or Slack via the Data agent or other plugins:

• Half-finished queries may leave ambiguous audit state — log what was asked even when answers fail
• Retry storms after recovery can spike warehouse costs — throttle reconnects
• Users may paste sensitive context into consumer chat as a workaround when Work is down — a shadow-IT risk worth addressing in policy

Security teams should treat Work availability like any other SaaS agent connector: incident response playbook, not just status-page watching.

What people are asking

Was ChatGPT completely down? No — the official affected component was ChatGPT Work under the ChatGPT product tree, not the entire platform or the API for all models.

Could I still use GPT-5 in normal chat? Reporting focused on Work mode; standard chat was not the named failure surface. Always verify live — partial outages can spread.

Did Plus users get refunds? OpenAI did not announce billing credits for this incident as of September 1. Check current policy if outages affect your procurement review.

Should we cancel ChatGPT Work pilots? Not necessarily — but pilots should include fallback documentation and explicit SLA assumptions in the business case.

How is this different from API outages? API failures often break programmatic integrations with clear error codes. Work mode failures hit interactive, long-horizon tasks where users may not know whether to wait, refresh, or abandon — worse for trust even when duration is similar.

Comparison — hosted Work vs local agent harness

Neither column wins outright — most mature teams in 2026 use both, with Work for speed and local harnesses for anything that must finish.

Bottom line

August 31, 2026 was a five-hour reminder that ChatGPT Work — OpenAI's flagship agentic surface for paid users — can go partially dark on a Monday afternoon while the rest of the AI news cycle moves on. The fix landed; the pattern remains. Builders routing production work through hosted Work threads need status monitoring, tier-aware impact models, and a second harness — or accept that agentic workflows will occasionally pause half a business day.

Related on explainx.ai

• OpenAI Launches a Data Agent in ChatGPT Work — and What It Means for Data Security — the next major ChatGPT Work feature after this outage
• Why AI agents haven't gone mainstream
• OpenAI rogue agent and pacing update
• What is an agent harness?
• Loop engineering for coding agents
• Grok Outlook read-write connectors
• ChatGPT ads $1B revenue run rate
• Build personal AI system local workflow

Official source: OpenAI status incident 01M1C5M4K0WC8PPT0Z175RJA1E.

Incident times and duration from OpenAI status logs as of September 1, 2026.