ChatGPT Goes Dark Globally as OpenAI Flags Login and API Failures

OpenAI’s ChatGPT went dark for users across the United States, India and several other regions on Friday morning, with the company’s status page confirming an active incident hitting conversations, account access and application programming interface (API) requests. Thousands of complaints piled onto Downdetector inside the first hour, with concentrations in New Delhi, Noida, Mumbai, Bengaluru and multiple US metros. ChatGPT now claims roughly 900 million weekly active users and an enterprise book OpenAI says clears more than 40% of company revenue, which means a one-hour stall sits in front of a customer base that includes 4 million developers writing against the same API.

What OpenAI’s Status Page Showed Friday

The OpenAI status page logged the disruption as a live incident covering three components: ChatGPT conversations, account login and account creation, and the developer API. Free chats and paid subscribers both hit errors mid-session. Sign-in pages bounced credentials. Developers with scheduled jobs against the same endpoints watched timeouts pile up.

A short note from the company said engineers had identified a problem and were applying fixes. Users on the consumer surface reported a mix of failure modes: chats that loaded but would not send, chats that hung mid-reply, and the familiar grey login screen that refuses to forward credentials.

OpenAI did not name a root cause in its first update. The incident touched both the free tier and paid Plus, Pro and Enterprise plans, judging by complaints across X, Reddit and OpenAI’s own developer forum. Codex, the company’s coding assistant, was caught in the same window because it shares authentication and session services with ChatGPT.

Flagging “conversations, account access and API” together points to a failure at the shared layer that authenticates users and routes requests. The blast radius covers the entire consumer chatbot, the developer console, and any application a third party has wired into ChatGPT as a back-end service.

Downdetector Caught the Spike Before the Fix

By the time OpenAI updated its status page, Downdetector’s ChatGPT incident tracker had already collected thousands of complaint pings. Reports jumped from a baseline of dozens per hour to thousands within minutes, peaked inside the first half hour, and tailed off as fixes rolled out. India logged complaints from the country’s main metros, with New Delhi, Mumbai, Bengaluru and Noida producing the loudest signal. US reports clustered around East Coast morning, suggesting the failure tracked traffic load rather than a regional fault.

The same shape hit on April 20, when more than 13,000 users reported issues at peak during a 90-minute partial outage that touched login, conversations and voice mode. February 3 and 4 set the high-water mark for the year so far, with more than 28,000 reports on day one and roughly 24,000 the next.

Two things matter about the Downdetector pattern. The platform is now the de facto early-warning system for ChatGPT, beating OpenAI’s status page by a few minutes on most events. And complaint counts no longer scale with real impact: against a user base in the hundreds of millions, even a 13,000-report spike translates to a tiny sliver of the affected population.

Why a One-Hour Stall Costs More Than It Did a Year Ago

A year ago, ChatGPT had 400 million weekly users and a much smaller paid base.

OpenAI now reports about 900 million weekly active users, more than double that figure. Annualized revenue passed $25 billion earlier this year, and the company has said it is on pace to cross a billion weekly users before year-end.

Reliance has gone deeper too. Enterprise customers contribute more than 40% of OpenAI’s revenue, the API processes 15 billion tokens per minute at peak, and the developer count sits at roughly 4 million.

When the consumer chatbot was the dominant surface, an outage meant some users opened Google instead. Today the same stall hangs internal copilots inside Fortune 500 deployments, breaks scheduled batch jobs, and freezes sales bots fielding customer queries at retail.

Recovery does not undo the damage. Inference pipelines that missed their window have to refill, user sessions that timed out get restarted from scratch, and customer service teams have to clear backlog before they can take new tickets.

The Enterprise Bill Behind a Consumer Hiccup

Inside enterprise IT departments, the same outage looks different from the consumer story. Microsoft Copilot uses OpenAI models under the hood for many features. Apple Intelligence routes some queries to ChatGPT under an existing partnership. Salesforce Einstein and Notion AI both keep OpenAI as a primary provider. Wyndham’s 8,400-hotel listing inside the ChatGPT app stops returning availability data when the consumer chat surface goes dark. When the shared session layer breaks, every one of those products loses a piece of its functionality during the window.

The corporate OpenAI enterprise tier sits behind contractual service-level agreements. SLAs (service-level agreements, the per-month uptime promises baked into enterprise contracts) typically allow credits past a 99.5% or 99.9% threshold, and a multi-hour outage like Friday’s eats straight into that buffer. By the numbers, here is what OpenAI now insures against on any given day:

• 900 million weekly active users on ChatGPT, double the figure from a year ago
• $25 billion in annualized revenue at the most recent disclosure
• 40%+ of company revenue from enterprise customers
• 4 million active developers writing against the OpenAI API

ChatGPT’s Outage Cadence Has Quietly Doubled

In the past 90 days, OpenAI’s status page has logged 59 incidents across its core services, according to StatusGator’s OpenAI outage tracker. One was a major outage and 58 were minor, with a median duration of 1 hour 47 minutes. That works out to one incident every 36 hours.

Over the same window, OpenAI’s twelve core components held about 99.83% uptime. The number sounds high but does not feel like it from a user seat, because the failures cluster. Two days in February alone produced more than 52,000 user reports between them. The April 20 event produced roughly 13,000 more. Most other days have run on smaller spikes that rarely cross 1,000 reports each.

The table below sets out the year’s noteworthy events to date:

The cadence carries implications beyond the headlines. A 90-day window with 59 incidents is roughly double the rate the same status page logged through most of 2025. Engineering load on OpenAI’s reliability team has gone up as a function of scale, model variety and the velocity of feature releases that include Sora, Codex updates and agent rollouts.

What Teams Did When the Cursor Stopped Blinking

Inside enterprise teams that depend on OpenAI’s API, the playbook for a live outage has become almost routine. It runs in three steps:

1. Route inference traffic to a secondary model provider on the same prompt schema
2. Surface a degraded-mode banner so end users see the failure tagged correctly
3. Queue non-urgent batch jobs until the primary service returns

Anthropic’s Claude, Google’s Gemini and a handful of open-weight models running on Together, Fireworks or local hardware show up as fallback options in most enterprise architecture diagrams now. Oton Technology covered the most common substitutions in a piece on the six AI tools enterprise teams reach for when ChatGPT and Codex are dark, which sets out the swap matrix many engineering shops keep on their incident-response page. Consumer users have fewer options: Claude.ai, gemini.google.com, Perplexity and Microsoft Copilot absorbed most of Friday morning’s bounce traffic, though none has been tested at ChatGPT’s user count.

The structural fix is the harder problem. Multi-model abstraction layers help a developer team stay up when a provider goes down, but they push complexity and cost onto the application layer. For a Fortune 500 customer signing a multi-year deal, the easier sell remains a single trusted vendor with a tighter SLA. Every outage like Friday’s chips at that pitch.

Frequently Asked Questions

Is ChatGPT Down Right Now?

OpenAI’s status page is the canonical source. The page at status.openai.com lists active incidents in real time and tags them by affected component. Downdetector usually registers a complaint spike a few minutes earlier than OpenAI’s official page during fresh events.

How Long Do ChatGPT Outages Usually Last?

The median incident duration over the past 90 days has been about 1 hour 47 minutes, according to OpenAI’s status data. Major outages have run up to 90 minutes for full restoration on the consumer surface, and minor ones often clear inside 30 minutes.

Does ChatGPT Going Down Affect Microsoft Copilot and Apple Intelligence?

Often, yes. Microsoft Copilot uses OpenAI models for several core features, and Apple Intelligence routes some queries to ChatGPT under a partnership. When OpenAI’s shared infrastructure breaks, those downstream products lose access to the affected models until the fix rolls out.

What Can Paid Plus and Enterprise Users Do During an Outage?

Refreshing the browser and clearing local cookies sometimes restores access during partial failures. Enterprise customers can fall back to Microsoft Azure OpenAI Service deployments if their contract includes that route. The corporate API publishes its own status feed for engineering teams running production workloads.