OpenAI stack hit concurrent ChatGPT, login, and access failures on 29 May 2026

Scope: This note covers vendor-reported failures on 29 May 2026 UTC. OpenAI had not published a detailed root-cause postmortem on the status history page at the time this incident note was drafted.

Facts below come from OpenAI Status history (May 2026 entries) and the live OpenAI Status dashboard. Secondary corroboration of user impact spikes appears in Downdetector and trade press summaries dated 29 May 2026; those sources are listed for context, not as primary technical diagnosis.

Verified facts

What happened

1. Early on 29 May, OpenAI logged separate active issues for ChatGPT conversations and for login or new account creation, both visible on the public history timeline.
2. The same morning, a distinct checkout incident affected business plan subscriptions on web and mobile web before recovery.
3. Later that day, OpenAI posted "Users May Experience Issues Accessing ChatGPT" with recovery noted on the history page.
4. May 2026 already showed elevated warning-level events on prior days (21, 26, 27, 28) in independent uptime trackers cited by secondary outlets; the 29 May cluster is the broadest same-day combination in that window.

PSF mapping

D8 Vendor resilience

Teams that route agents, batch jobs, and customer-facing chat through a single OpenAI API key without a tested fallback experienced hard failure when conversation and auth paths degraded together. See the D8 vendor resilience guide and the OpenAI Agents SDK PSF assessment for dependency trade-offs.

Practitioner action: Document a provider failover runbook with health checks on both chat completion and auth-dependent features, then rehearse it quarterly.

D4 Observability

External status pages lagged user pain on social channels in some regions. Production stacks should emit their own SLO dashboards (error rate, latency, queue depth) and alert before public status moves to degraded.

Practitioner action: Add synthetic probes against your critical model routes and map alert thresholds to customer-visible failure, not only HTTP 200 from a health endpoint.

D5 Deployment safety

Checkout and workspace flows failing alongside inference paths suggests shared control-plane dependencies. Deployments that assume "API up" equals "product up" need segmented degradation modes (read-only, cached responses, human handoff).

Practitioner action: Ship a feature flag bundle that disables non-essential agent tools while keeping audited fallback answers for tier-1 user journeys.

Prevention controls

• Maintain a secondary model route (different vendor or region) with automated circuit breaking when OpenAI error budgets exceed policy.
• Cache last-known-good prompts and retrieval contexts for read-heavy agent workflows during outages.
• Publish customer comms templates tied to status.openai.com component names your product actually uses.
• Track vendor incident IDs in your internal incident registry for post-incident reviews and SLA credit requests.

Production context

This event is distinct from single-model latency spikes: multiple user-facing surfaces failed on the same day. For taxonomy and monitoring patterns, see seven failure modes of production AI, PSF Domain 8, and the production-ready checklist.

Sources

• OpenAI Status: History (May 29, 2026 entries)
• OpenAI Status: Current system status
• Downdetector: OpenAI (user reports, 29 May 2026) (secondary)
• News9live: OpenAI service disruption report (29 May 2026, secondary)
• AI Weekly alert summary (29 May 2026, secondary)