Key Takeaways:<\/p>\n
AI app disruption stepped up sharply in Q1 2026. Across ChatGPT, Claude, Gemini, and Copilot, high-signal disruption days rose from six in Q1 2025 and 16 in Q4 2025 to 51 in Q1 2026. Claude accounted for 39 of those 51 service-days, while Gemini accounted for seven, Copilot three, and ChatGPT two.<\/p>\n
OpenAI\u2019s ChatGPT produced the largest individual AI app disruption signals, but its baseline trend has improved substantially. ChatGPT accounted for four of the five largest AI app days in our data, including roughly 68,000 reports on December 2, 2025. Yet its daily median report volume was lower in April 2026 than in April 2025, pointing to improving reliability over time even as Codex usage has scaled rapidly in recent months.<\/p>\n
Claude became the clearest example of scale-up volatility. Claude recorded near-zero Downdetector report volumes in early 2025, then moved into a sustained report baseline from mid-July as adoption rose. By Q1 2026, it accounted for 39 high-signal AI app disruption days, with March report volume nearly three times February\u2019s level alone.<\/p>\n
Hyperscaler incidents are part of the AI reliability surface. Cloud infrastructure like AWS and Microsoft Azure can create a massive blast radius when problems arise, explaining part of the AI risk surface. AWS\u2019s October 20, 2025 DynamoDB event<\/a> and Azure\u2019s October 29, 2025 Front Door event<\/a> were infrastructure-layer shocks that showed how failures in cloud control planes can quickly become visible to AI users.<\/p>\n \n\t\t\tSevere AI outage days are rising AI reliability now spans multiple failure layers<\/p>\n AI platforms are not single systems from the user\u2019s point of view, even when they present a single interface. A ChatGPT, Claude, Gemini, or Copilot failure can sit in the product layer, the provider orchestration layer, the hyperscaler layer, or the edge and access layer.<\/p>\n The product layer is what users actually see. The provider orchestration layer includes login, routing, model selection, rate limits, feature flags, inference scheduling, retry behavior, and capacity allocation. The hyperscaler layer includes compute, databases, storage, networking, and regional control planes. The edge and access layer includes DNS, web gateways, bot protection, content delivery, and authentication flows.<\/p>\n Those layers are not always owned by different companies, and they are not the full physical internet stack. Network operators, subsea cables, data centers, and user access networks still matter. The focus here is narrower: the service and dependency layers that are most visible in Downdetector data and public incident records.<\/p>\n This distinction is important because the same user-facing symptom can have different operational meanings. A failed prompt, login loop, missing chat history, rate-limit error, unavailable file, or stalled agent task may not share the same root cause. For enterprise buyers and risk teams, resilience is about understanding more than whether an AI platform was simply available. They need to know where the issue occurred, which workflows were affected, and whether it reflected a problem with a single provider or a broader dependency across the AI stack.<\/p>\n OpenAI sees ChatGPT and Codex reliability improve over time<\/p>\n OpenAI dominated the largest raw AI app disruption days. Its top Downdetector days were December 2, 2025 with 67,567 reports, February 4, 2026 with 55,039, July 16, 2025 with 51,420, June 10, 2025 with 44,517, and February 6, 2025 with 41,674. Each was far above OpenAI\u2019s full-period median of 1,428 reports.<\/p>\n Most days look more stable than those headline events suggest. From January 14, 2025 onward, OpenAI\u2019s underlying log trend moved modestly downward rather than upward, with an annualized multiplier of 0.76x. Its monthly median daily report volume peaked in April 2025 at 2,157 and was 1,166 in April 2026. That points to an improving report baseline over time, even as OpenAI said ChatGPT had more than 900 million weekly active users<\/a> by early 2026 and later said more than 4 million people used Codex<\/a> each week.<\/p>\n Since OpenAI\u2019s services are operating at enormous scale, and failure events become highly visible because ChatGPT and Codex are increasingly embedded in work, search, coding, and agentic tasks, it makes natural sense that its outage volumes are higher nominally in our data. OpenAI said<\/a> daily Codex usage grew by more than 10x from early August to October 2025, and GPT-5-Codex served more than 40 trillion tokens in its first three weeks.<\/p>\n The public incident record also points to mixed causes rather than one uniform dependency. On June 10, 2025<\/a>, OpenAI attributed elevated errors to a host operating system update on cloud-hosted GPU servers that caused many GPU nodes to lose network connectivity. ChatGPT error rates reached about 35% at peak and API error rates about 25%. On February 3, 2026<\/a>, OpenAI described a configuration change that introduced an unexpected data type in a critical execution path, with retry traffic amplifying downstream load and slowing recovery in one region. By June 2026, OpenAI said<\/a> Codex had surpassed 5 million weekly users, up more than 6x since the February desktop app launch.<\/p>\n Taken together, those events point to provider orchestration and capacity systems rather than one simple cloud dependency. OpenAI\u2019s resilience picture, like other AI platforms, spans GPU fleet management, configuration validation, retry behavior, login and conversation availability, and web access.<\/p>\n Claude becomes the volatility story as adoption and capacity scale<\/p>\n Anthropic\u2019s outage profile is notably different. Claude recorded near-zero Downdetector reports in early 2025, which means there was little sustained user-reported signal in our data, not that there were necessarily no issues. The service then moved into a clearer report baseline in mid-July, reflecting its later rise in popularity compared with ChatGPT. After that, report volume accelerated sharply. Claude generated 48,589 reports in Q3 2025, 108,694 in Q4 2025, and 314,996 in Q1 2026. Its median monthly daily reports rose from 285 in July 2025 to 2,830 in March 2026.<\/p>\n The upper end of Claude\u2019s report distribution rose with that larger baseline. Claude had four high-signal disruption days in Q3 2025, 12 in Q4 2025, and 39 in Q1 2026. Its March 2026 total of 192,773 reports was almost three times February\u2019s total and more than five times December\u2019s total. That does not necessarily mean Claude became less reliable in a simple linear way, because the platform was moving into a different scale of usage, where more people were using it more often and heavier workloads with Claude Code and Cowork became more visible when things went wrong.<\/p>\n The causes appear mixed, which is an important point. Claude\u2019s Q1 2026 pattern cannot be reduced to one outage. Downdetector spikes clustered around demand surges, model-release windows, and platform instability. Anthropic\u2019s March 2026 usage promotion<\/a> doubled usage limits outside weekday peak hours for eligible users, which is consistent with a platform trying to shape demand when capacity is under pressure.<\/p>\n Claude\u2019s largest day, 46,676 reports on April 15, 2026, came one day before the Opus 4.7 release<\/a>, while elevated report activity also appeared around earlier Opus release windows, including Opus 4.5<\/a>. This does not prove model launches caused the outages, but it does show that frontier model shipping, demand management, and user-visible reliability are now tightly coupled.<\/p>\n Anthropic\u2019s own public messaging underscores the growth side of the story. In February 2026, it said Claude Code had passed $2.5 billion in run-rate revenue, with weekly active users doubling and business subscriptions quadrupling since January. By late May, Anthropic said total run-rate revenue had reached $47 billion, up from $14 billion in February, while raising $65 billion<\/a> at a $965 billion valuation. Its SpaceX capacity deal<\/a>, adding access to 300+ MW and 220,000+ NVIDIA GPUs, also supported higher Claude Code and Opus limits, reinforcing the link between demand growth, compute supply and platform-scale monetisation.<\/p>\n \n\t\t\tMonthly Median Daily Reports by AI Service Gemini and Copilot show smaller but distinct outage patterns<\/p>\n Google\u2019s Gemini showed the clearest upward shift among services, excluding Claude, where our data covers the full baseline period. Its report baseline rose from January 2025 onward, while high-signal disruption days increased from zero in Q1 2025 to seven in Q1 2026.<\/p>\n This fits the wider growth story around Gemini. Google said the Gemini app had reached more than 750 million monthly active users<\/a> by Q4 2025, and later said more than 900 million people<\/a> used Gemini monthly by May 2026, so higher report volumes are partly a function of a much larger user base.<\/p>\n Gemini\u2019s largest outage day in our data was February 13, 2026, with 14,417 reports, six days before Google\u2019s Gemini 3.1 Pro<\/a> announcement on February 19. The timing supports a cautious launch-window link (the data alone does not prove causality), especially given the phased rollout across the Gemini app, Gemini API, Vertex AI, and NotebookLM.<\/p>\n \n\t\t\tQuarterly AI App Report Volumes by Service Copilot\u2019s outage reports are smaller in absolute volume, but its outage pattern looks more tied to the Microsoft ecosystem. Its largest day was June 4, 2025 with 12,028 reports, about 81 times its full-period median of 149. Copilot also co-spiked with OpenAI on six same-day events in our separate co-spike analysis, the strongest AI app pairing in that matrix. We also observed much less outage reports about Copilot on weekends, reflecting the enterprise-aligned use of the service.<\/p>\n This pattern is consistent with Copilot being less of a standalone frontier-model destination and more of an AI layer inside Microsoft\u2019s wider productivity, identity, and enterprise data stack. Microsoft\u2019s own architecture describes Copilot as grounding prompts through Microsoft Graph and the Semantic Index<\/a>, while respecting user identity and access boundaries, which means failures can surface through several layers, not just the model itself.<\/p>\n Hyperscalers matter to AI resilience<\/p>\n The cloud layer is important because AI services are built on top of it. AWS and Azure are not AI applications in this analysis. They are hyperscalers, meaning the large cloud providers that supply compute, storage, databases, networking, login systems, and traffic routing that many digital services rely on. Their outage reports should not be counted as AI app outages. They are critical because they show stress in infrastructure underneath AI platforms and enterprise AI workflows.<\/p>\n AWS\u2019s October 20, 2025 event<\/a> was the largest infrastructure shock in our time series, with 315,342 U.S. Downdetector reports for AWS. AWS traced the main issue to a latent race condition in DynamoDB\u2019s DNS management system that produced an empty DNS record for a key regional endpoint in us-east-1. In plain terms, an automated system briefly failed to tell dependent services where to connect. The recovery then became more complicated, with follow-on issues affecting EC2 instance launches and Network Load Balancers.<\/p>\n Azure\u2019s October 29, 2025 Front Door incident produced 95,840 Azure reports. Front Door is Microsoft\u2019s global traffic entry and routing service, so problems there can affect how users reach applications even when the application itself is not the root cause. Microsoft\u2019s lessons-learned post<\/a> said incompatible customer configuration metadata moved through protections before a delayed processing task crashed. Microsoft emphasized stronger validation and control-plane hardening.<\/p>\n These events should not be used to explain every AI platform spike, and they are not evidence that hyperscalers are uniquely fragile. AWS and Microsoft both published detailed post-event analysis and remediation steps, which is exactly how large-scale infrastructure providers improve resilience. The point is that when AI workflows depend on cloud databases, traffic routing, identity systems, edge networks, and management software, failures in those layers become part of the AI reliability surface.<\/p>\n Microsoft\u2019s October 2025 OpenAI partnership update<\/a> reinforces that this dependency map is changing and not disappearing. OpenAI can serve non-API products on any cloud provider, while API products developed with third parties remain exclusive to Azure, and OpenAI has contracted to purchase an additional $250 billion of Azure services. Multi-cloud optionality can improve flexibility, but enterprise resilience still depends on knowing which products, APIs, regions, and workflows sit on which dependencies.<\/p>\n AI reliability is moving beyond chat availability<\/p>\n Evidence that surfaced after our April webinar<\/a> on this topic strengthens the same points. As AI tools move into coding, file handling, APIs, connectors, and enterprise workflows, the weak points become more varied. A service can appear broadly available while a specific model, agent feature, upload path, integration, or developer API is failing.<\/p>\n OpenAI\u2019s public status history<\/a> through late May listed repeated incidents outside generic ChatGPT availability, including Responses API, Codex, Code Review, Realtime API, file uploads, workspace connectors, image generation, transcription, and model-specific errors. A May 8 Responses API incident<\/a> was linked to a recent deploy and rollback, while a May 13 Codex 5.5 incident<\/a> was model-specific.<\/p>\n
Days exceeding 10x each service’s own full-period daily median \u2014 Downdetector U.S., Jan 2025-Apr 2026<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/06\/AI-Outage-Layers-1200x675.png\")
<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/06\/OpenAI-Reliability-Improvement-1200x675.png\")
<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/06\/Claude-Outages-1200x675.png\")
<\/p>\n
Typical daily Downdetector report volume \u2014 U.S., Jan 2025-Apr 2026<\/p>\n
Total Downdetector user-reported problems per quarter \u2014 U.S., Q1 2025-Q1 2026<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/06\/Hyperscaler-Disruptions-1200x675.png\")
<\/p>\n![\"\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/06\/CDN-Outage-Importance-1200x675.png\")
<\/p>\n