OpenAI Outage History: A Record 300 Incidents in 2026, but Full Outages Fell to 6
How often does ChatGPT go down? OpenAI’s status page logged 300 incidents from January 1 to October 29, 2026, more than in any full prior year, and 199 of them touched ChatGPT. Most were partial slowdowns: only 6 reached a full outage on any component, down from 14 in the same months of 2025. This OpenAI outage history counts all 1,051 incidents since 2021.
We downloaded every incident record from the JSON feed behind status.openai.com, from the first entry on February 12, 2021 to October 27, 2026, plus the 75 post-incident write-ups OpenAI attached to them. The count of incidents keeps climbing as OpenAI adds products and reports problems at a finer grain. The severe ones have moved the other way.
Every OpenAI incident-day since 2024
Worst status that day: degraded performance partial outage full outage no component marked no incident
One square per day (UTC), coloured by the worst component status of any incident that began that day. Hover or tap a square for the count. On a phone, swipe the calendar sideways.
In 2024 there was at least one incident on 130 days. In 2025 that rose to 207 days, and 2026 has had 167 incident-days in its first 302 days. The red squares, days with a full outage, have thinned out: 16 full-outage incidents in 2024, 16 in 2025, and 6 so far in 2026. Incident-days actually fell this year, 167 against 177 by the same date in 2025, even as the incident count rose: problems now cluster, with several narrow incidents logged on the same day.
How often does ChatGPT go down?
Counting status-page incidents that list a ChatGPT component, ChatGPT had 93 incidents in 2024, 221 in 2025 and 199 from January 1 to October 29, 2026. That is roughly 20 a month this year. Most are scoped to one feature, plan or region, so a single user is usually unaffected. A more useful measure is time in outage: we merged overlapping incidents and added up the hours each core component spent in partial or full outage.
| Jan 1 to Oct 29 | Incidents | Days with an incident | Reached partial outage or worse | Reached full outage | Median duration | ChatGPT Conversations, outage hours | Chat Completions API, outage hours |
|---|---|---|---|---|---|---|---|
| 2023 | 146 | 102 | 69 | 30 | 1 h 7 min | 26.8 | 60.9 |
| 2024 | 164 | 109 | 54 | 12 | 1 h 27 min | 72.1 | 45.8 |
| 2025 | 284 | 177 | 66 | 14 | 2 h 1 min | 27.7 | 23.7 |
| 2026 | 300 | 167 | 39 | 6 | 1 h 44 min | 16.5 | 0.9 |
Same calendar window each year, so 2026 compares like for like. Outage hours are the merged time a component spent in partial or full outage; degraded performance is excluded. Median duration runs from the first sign of impact on the status page to the first “resolved” update.
ChatGPT Conversations, the component behind sending and receiving messages, spent 16.5 hours in partial or full outage this year to October 29, against 27.7 hours in the same window of 2025 and 72.1 hours in 2024. It has not been marked as fully down at all in 2026, against 1.5 hours in 2025. For developers, the Chat Completions API logged 0.9 hours of partial or full outage, down from 23.7 hours a year earlier and 60.9 hours in 2023. ChatGPT’s growth, tracked in our ChatGPT market share report, makes those hours more costly each year even as they shrink.
Incidents per month
degraded partial outage full outage none marked
Incidents by the month they began, stacked by worst status. October 2026 runs to October 29.
The busiest month on record is August 2025, with 41 incidents. The step up in early 2025 lines up with OpenAI’s move to a new status page in February 2025, which splits products into many more components, and with new products getting their own entries: Codex from May 2025, then FedRAMP, ChatGPT Work and the Ads Platform in 2026. More reporting surface means more incidents on the page without a matching rise in severe ones.
Full years
| Year | Incidents | Days with an incident | Partial or worse | Full outage | Median duration | ChatGPT incidents | API incidents | Conversations outage hours | Chat Completions outage hours |
|---|---|---|---|---|---|---|---|---|---|
| 2021 | 10 | 9 | 3 | 3 | 2 h 41 min | 0 | 5 | 0.0 | 1.2 |
| 2022 | 43 | 39 | 24 | 3 | 47 min | 0 | 35 | 0.0 | 19.0 |
| 2023 | 172 | 119 | 81 | 36 | 1 h 9 min | 52 | 95 | 34.9 | 72.0 |
| 2024 | 187 | 130 | 64 | 16 | 1 h 28 min | 93 | 74 | 91.2 | 51.3 |
| 2025 | 339 | 207 | 79 | 16 | 2 h 3 min | 221 | 118 | 30.3 | 23.7 |
| 2026* | 300 | 167 | 39 | 6 | 1 h 44 min | 199 | 74 | 16.5 | 0.9 |
*2026 runs to October 29. 2021 and 2022 entries were sparse and mostly backfilled. The record full year before 2026 was 339 incidents (2025).
Which OpenAI products have the most incidents
Since March 2025, when component-level reporting began, the status page has logged 596 incidents in 20 months. ChatGPT appears in most of them; the API in about a third. Sora has never had its own status component, so we count it by name: 25 incident titles mention Sora since 2024.
| Product group (Mar 2025 to Oct 29, 2026) | Incidents | Share | Partial or worse | Full outage | Median duration |
|---|---|---|---|---|---|
| ChatGPT | 394 | 66% | 73 | 7 | 1 h 54 min |
| APIs | 177 | 30% | 38 | 6 | 1 h 53 min |
| Codex | 97 | 16% | 23 | 3 | 1 h 46 min |
| FedRAMP | 10 | 2% | 2 | 1 | 5 h 44 min |
| Ads Platform | 5 | 1% | 1 | 0 | 53 min |
| Sora (named in title) | 20 | 3% | 8 | 4 | 1 h 25 min |
| No component listed | 65 | 11% | 8 | 5 | 1 h 54 min |
One incident can list several groups, so shares add to more than 100%. ChatGPT Atlas is counted with ChatGPT.
The components most often listed in that period:
- Conversations 189
- Login (ChatGPT) 117
- Codex Web 94
- Responses 88
- Image Generation 80
- File uploads 76
- Chat Completions 72
- Search 71
- Voice mode 68
- Compliance API 67
- Agent 63
- Images 61
Degraded performance vs partial vs full outage
OpenAI marks each affected component with one of three statuses. Degraded performance means slower responses or elevated errors; partial outage means a large share of requests failing; full outage means the component is unavailable. We label each incident by the worst status it reached.
| Worst status (Mar 2025 to Oct 29, 2026) | Incidents | Share | Median duration |
|---|---|---|---|
| Degraded performance | 444 | 74% | 1 h 50 min |
| Partial outage | 86 | 14% | 2 h 17 min |
| Full outage | 18 | 3% | 1 h 5 min |
| No component marked | 48 | 8% | n/a |
Across all 585 incidents in that period with a usable duration, the median ran 1 h 54 min from first impact to the “resolved” update, and the 90th percentile 15 h 15 min. 30% closed within an hour and 30% stayed open more than four hours. The long tail is mostly narrow problems, such as delayed Compliance API data or one region’s Deep Research errors, not broad outages. The longest incident in the whole record, 21 d 5 h 5 min, was “New users may experience temporary delays in video generation capabilities” in April 2025.
When OpenAI outages happen
Starting hour of 596 incidents, March 2025 to October 29, 2026. ET follows daylight saving time.
93% of incidents begin on a weekday (555 of 596); Saturdays and Sundays together account for 41. The busiest hour is 13:00 to 13:59 Eastern, and 43% of incidents start between 10 a.m. and 4 p.m. Eastern, which is 7 a.m. to 1 p.m. in San Francisco. The pattern holds for severe incidents too: 99 of 104 partial or full outages began on a weekday.
The data cannot separate the two obvious explanations, weekday traffic peaks and weekday engineering changes, because both follow the same clock. What OpenAI’s write-ups do show is that most explained incidents began with a change OpenAI made, and changes ship on working days.
Incidents by weekday (US Eastern)
The biggest OpenAI outages and what caused them
These are the largest incidents by reach and length that come with an OpenAI write-up. “Status window” is how long the status-page incident ran; OpenAI’s write-ups sometimes give a shorter impact window.
| Date (UTC) | Incident (links to OpenAI write-up) | Worst status | Status window | Cause stated by OpenAI (our summary) | Cause type |
|---|---|---|---|---|---|
| 2026-09-29 | Elevated errors across ChatGPT, Codex, and the API including the Agents API | Degraded performance | 5 h 21 min | Feature rollout exhausted internal network connections | OpenAI’s own change |
| 2026-07-23 | Elevated Error Rates | Degraded performance | 1 d 3 h 57 min | Provider network maintenance failure cut a region | Cloud or third-party provider |
| 2026-06-03 | Elevated error rates on Codex, ChatGPT and Responses API | Partial outage | 6 h 11 min | Configuration rollout exhausted a shared backend | OpenAI’s own change |
| 2026-04-20 | Users unable to load ChatGPT, Codex and API Platform | Partial outage | 4 h 13 min | Authentication edge dependency overloaded | Internal system failure |
| 2025-11-18 | Access issues affecting OpenAI websites | Full outage | 3 h 6 min | Faulty config rollout at an upstream networking provider | Cloud or third-party provider |
| 2025-06-10 | Elevated error rates | Partial outage | 15 h 24 min | Routine host OS update cut network on GPU nodes | OpenAI’s own change |
| 2025-03-30 | Issues with new sign ups | Degraded performance | 1 d 12 h 53 min | Signup surge after a feature launch saturated account services | Demand or capacity |
| 2025-01-23 | Increased errors for ChatGPT | Full outage | 2 h 52 min | Provider Cosmos DB failure, then crash-looping web pods | Cloud or third-party provider |
| 2024-12-26 | High error rates for ChatGPT, APIs, and Sora | Full outage | 9 h 38 min | Power failure in a cloud provider data center | Cloud or third-party provider |
| 2024-12-11 | API, ChatGPT & Sora Facing Issues | Full outage | 7 h 6 min | New telemetry service overwhelmed Kubernetes control planes | OpenAI’s own change |
| 2024-06-04 | ChatGPT is unavailable for some users. | Full outage | 2 h 44 min | ChatGPT database became unreachable under pooled connection surges | Internal system failure |
| 2023-11-08 | Major Outage across ChatGPT and API | Full outage | 1 h 51 min | Record request volume pushed routing nodes past memory limits | Demand or capacity |
| 2023-02-21 | Outage on all models | Full outage | 4 h 30 min | Cloud provider database maintenance left read replicas down | Cloud or third-party provider |
Since March 2025 the widest incident by component count was “Degraded performance in ChatGPT, Codex, and APIs” on October 21, 2026, which listed 33 components across APIs, ChatGPT and Codex, though at degraded performance.
Not every big outage starts inside OpenAI. On November 18, 2025, OpenAI’s websites returned 403 and 504 errors from 3:30 to 6:40 AM PST, which its write-up blames on “an upstream third-party networking provider” without naming it. That window, 11:30 to 14:40 UTC, falls inside the Cloudflare outage that began at 11:20 UTC the same morning. API traffic and the mobile apps were unaffected, per OpenAI. On December 26, 2024, a power failure in a cloud provider’s data center pushed error rates above 90% for ChatGPT, Sora and many APIs for hours, because region-wide database failover “currently requires manual intervention from the hosting cloud provider”.
What causes OpenAI outages
OpenAI has attached a write-up to 79 status incidents, which describe 75 distinct events. We read each one and coded the trigger it names. 48 of 75 (64%) began with OpenAI’s own deployment, configuration change, code change, migration or experiment. 11 trace to a cloud or third-party provider (data center power, provider maintenance, an authentication vendor, a networking provider). 9 were internal system failures with no change named, mostly databases and caches, and 7 were demand outrunning capacity.
OpenAI’s own change demand or capacity provider internal failure
11 of the 18 write-ups for incidents that reached full outage name an OpenAI change. Since 2025, 14 of the 28 own-change write-ups name a configuration change specifically, and 19 of all 43 write-ups from that period mention login, sign-in or authentication failures. Write-ups are not published for every incident: 10 of the 39 incidents that reached partial or full outage in 2026 have one, so this is a sample of the incidents OpenAI chose to explain, weighted toward the larger ones.
Is OpenAI getting more reliable?
On severity, yes. Comparing January 1 to October 29 each year, incidents that reached partial or full outage fell from 66 in 2025 to 39 in 2026, full outages from 14 to 6, and hours of partial or full outage on the Chat Completions API from 23.7 to 0.9. On frequency, no: total incidents rose from 284 to 300, though they were spread over fewer days (167 against 177) and the median incident was shorter, 1 h 44 min against 2 h 1 min. Two caveats apply. The 2025 status-page migration changed how incidents are recorded, so counts before and after March 2025 are not strictly comparable. And status-page labels are OpenAI’s own; the page footer says availability is “reported at an aggregate level across all tiers, models and error types”, so one customer’s experience can be worse than the page shows.
Search every OpenAI incident
All 1,051 incidents from the status page. Search titles, filter by year, worst status or product, and sort by date or duration. Each title links to the incident on status.openai.com.
| Start (UTC) | Incident | Products | Worst status | Duration |
|---|
What this means for MSPs
Set expectations in hours, not incidents. A client who reads “300 OpenAI incidents this year” will overestimate the risk. The better number for a business that runs on ChatGPT is the 16.5 hours ChatGPT Conversations spent in partial or full outage from January to October 2026, plus the much longer stretches of degraded performance. Put that next to the client’s own cost of downtime before deciding how much redundancy to buy.
Plan for Tuesday at 1 p.m., not Sunday night. 93% of incidents start on a weekday and the peak is early afternoon Eastern, the middle of the US working day. Any workflow that cannot wait an hour (customer chat, intake, code review gates) needs a manual fallback that staff have practiced.
Fail over by feature, not by vendor. 56% of incidents since March 2025 list a single component, and 72% list three or fewer: login, file uploads, a model, Codex, voice. Applications built on the API can route to a second model or a second provider when error rates climb, and they should treat a login outage separately, because 19 of OpenAI’s 43 write-ups since January 2025 mention login, sign-in or authentication failures. Test the fallback path regularly; an untested second provider is not a failover.
Monitor the feed, not social media. OpenAI’s status page exposes an RSS feed (status.openai.com/feed.rss), an Atom feed (status.openai.com/history.atom) and Statuspage-style JSON (status.openai.com/api/v2/summary.json, /api/v2/incidents.json). Wire one into the RMM or PSA so a ticket opens before the client calls. The status page can lag the impact: in 82 of 596 incidents since March 2025, the first component impact was backdated more than five minutes before the incident was posted.
Read the SLA before promising one. OpenAI’s public terms put a 99.9% uptime SLA only on Scale Tier, sold to Enterprise API customers, and on Fast mode traffic for customers on enterprise agreements, with service credits as the remedy. Standard API requests are excluded, and the Services Agreement provides the services “AS IS”. We found no public uptime commitment for ChatGPT subscriptions. Our 99.9% uptime explainer covers what that level allows: about 44 minutes of downtime in an average month.
Is ChatGPT down right now? How to check
This report is a history; it does not update live. For the current state, open status.openai.com, which shows each ChatGPT, API, Codex, FedRAMP and Ads component. If the page shows everything operational but ChatGPT fails for you, OpenAI’s own footer notes that individual availability can differ by subscription tier, model and feature. Check whether the problem is limited to one feature (voice, file uploads, login) and try the mobile app, which stayed up during the November 18, 2025 web outage.
Frequently asked questions
How often does ChatGPT go down?
OpenAI’s status page logged 199 incidents touching ChatGPT components between January 1 and October 29, 2026, about 20 a month, out of 300 incidents across all products. Most were partial slowdowns or errors for some users. ChatGPT Conversations spent 16.5 hours in partial or full outage over that period, against 27.7 hours in the same months of 2025.
Is ChatGPT down right now?
This page is a history, not a live monitor. Check status.openai.com, or subscribe to its RSS feed (status.openai.com/feed.rss). Automated tools can poll status.openai.com/api/v2/status.json and /api/v2/incidents.json, which return the current state and the latest 25 incidents.
What was the biggest OpenAI outage?
By OpenAI’s own account, December 11, 2024: a new telemetry service overwhelmed the Kubernetes control plane in most large clusters, and all OpenAI services were significantly degraded or unavailable from 3:16 PM to 7:38 PM PST (4 h 22 min). The status-page incident stayed open 7 h 6 min. The longest broad partial outage was June 10, 2025, at 15 h 24 min, after a routine OS update cut network connectivity on GPU servers.
What causes OpenAI outages?
OpenAI has published 75 distinct post-incident write-ups on its status page. By our coding, 48 (64%) trace to OpenAI’s own deployment, configuration or code change, 11 to a cloud or third-party provider, 9 to an internal system failure such as a database, and 7 to demand outrunning capacity.
Is OpenAI getting more reliable?
Incident counts are up: 300 from January 1 to October 29, 2026, against 284 in the same period of 2025 and 164 in 2024. Severe incidents are down: 39 reached partial or full outage, against 66 in 2025, and only 6 reached full outage, against 14. The Chat Completions API spent 0.9 hours in partial or full outage, against 23.7 hours.
Does OpenAI have an uptime SLA?
OpenAI’s public Scale Tier page offers a 99.9% uptime SLA for Enterprise API customers who buy reserved capacity, and its Fast mode page (formerly Priority processing) says those SLAs are treated the same, with service credits for customers on enterprise agreements. Standard API requests are not eligible, and the public Services Agreement provides the services as is. We found no public uptime commitment for ChatGPT plans.
Methodology
Source. OpenAI status page (status.openai.com, run on incident.io). We pulled every incident through the JSON endpoint the page itself uses, /proxy/status.openai.com/incidents, one calendar month at a time with its start_at and end_at parameters, splitting any window that hit the 100-record page size, and removed duplicates where an incident spanned two months. Retrieved October 30, 2026. The pull holds 1,052 records: 1,051 incidents and 1 maintenance entry (excluded), from February 12, 2021 to October 27, 2026.
Completeness checks. Every incident in the endpoint’s default list (latest 100) and in the Statuspage-compatible /api/v2/incidents.json (latest 25) is in our pull, as is every incident linked from the rendered history page. For the period before February 2025, we compared Internet Archive captures of the old status page’s history view for eight complete months (May, June, October and November 2023; April, May, October and November 2024): the 107 incidents listed there match our records one for one by month and title.
Definitions. An incident is one status-page incident, whatever the number of components it lists; overlapping incidents are counted separately. Start is the earliest of the publish time and the first component impact; end is the first update marked “resolved”, so later write-up postings do not stretch durations. 42 incidents, mostly 2021 to 2023 entries posted after they were resolved, have no usable duration and are left out of duration statistics. Worst status is the highest of degraded performance, partial outage and full outage across the incident’s component impacts and status summaries. Outage hours merge overlapping partial and full outage intervals on one component. “Days with an incident” uses the UTC start date. Same-period comparisons use January 1 to October 29 of each year.
Products. Product groups come from the status page’s current component tree (APIs, ChatGPT, Codex, FedRAMP, Ads Platform). Incidents published before February 25, 2025 were imported from OpenAI’s previous Statuspage site, where one ChatGPT or API component was expanded to several on import; we therefore report components only from March 2025, and groups for all years. Sora has no component, so Sora counts use incident titles.
Cause coding. 79 incidents carry an OpenAI write-up. Where several incidents share one write-up (the March 2023 DALL-E events and July 25, 2026) we count one event, giving 75. Each event was coded by The Uptime Report to the first trigger the write-up names; a key sentence from each is checked verbatim against the source text when this page is built. Coding is our judgement, and write-ups exist for a minority of incidents.
Limitations. Status-page incidents are OpenAI’s own reports; they do not measure what any one customer saw, and OpenAI states availability is aggregated across tiers, models and error types. New products add components over time, and the 2025 migration changed recording granularity, so long-run counts mix reporting changes with reliability changes. Time-of-day figures use the incident start time. Counts include every status-page incident, including a small number of notices about customer support, billing or admin consoles that list no service component (37 incidents in 2026 list no component).
Not advice. This is an analysis of public status data, not a recommendation about any vendor or contract.
Cite as:The Uptime Report. "OpenAI Outage History: A Record 300 Incidents in 2026, but Full Outages Fell to 6." October 2026. https://theuptimereport.com/openai-outage-history/
References
- OpenAI Status, incident history and component tree. Accessed October 2026.
- OpenAI Status, incidents JSON endpoint (queried by month with start_at and end_at). Accessed October 2026.
- OpenAI Status, Statuspage-compatible API v2. Accessed October 2026.
- OpenAI, write-up: API, ChatGPT and Sora Facing Issues, December 11, 2024. Accessed October 2026.
- OpenAI, write-up: Elevated error rates, June 10, 2025. Accessed October 2026.
- OpenAI, write-up: High error rates for ChatGPT, APIs, and Sora, December 26, 2024. Accessed October 2026.
- OpenAI, write-up: Access issues affecting OpenAI websites, November 18, 2025. Accessed October 2026.
- Cloudflare, Cloudflare outage on November 18, 2025. Accessed October 2026.
- Internet Archive, captures of status.openai.com/history, 2023 to 2025. Accessed October 2026.
- OpenAI, Scale Tier for API Customers. Accessed October 2026.
- OpenAI, Fast mode for API Customers (formerly Priority processing). Accessed October 2026.
- OpenAI, Services Agreement (effective January 1, 2026). Accessed October 2026.
Leave a Reply