Moksh Gupta

Posted on Oct 13 Originally published at devtoollab.com

LLM API Outages in 2026: OpenAI and Claude Overlapped for 36 Minutes in 90 Days

#ai #programming #devops #webdev

If your product calls a hosted model, its worst days are partly your provider's worst days. The only public record of those days is a status page, and every provider runs its own, with its own rules for what gets declared. Most of the numbers people pass around are also old: an IsDown report from December 2025, which Vercel still cites in its AI gateway guide, counted 22 OpenAI incidents and roughly 182.7 hours of incident time in one month.

So we collected fresh data: every incident posted on 12 LLM provider status pages between July 14 and October 11, 2026. OpenAI's model APIs spent 72.7 hours with an incident open. Anthropic's Claude API spent 76.5. The two were in an incident at the same time for only 36 minutes, and both times the two sides were broken in different ways. The full write-up, with the per-provider tables and a free CSV of all 459 incidents, is on DevToolLab.

The Short Version

OpenAI posted 109 incidents across all its products in the window. Twenty touched its model APIs, which had an incident open 3.4% of the 2,160 hours.
Anthropic posted 82. Sixty touched the Claude API, open 3.5% of the time. Its median API incident ran 51 minutes; OpenAI's ran 124.
The two APIs overlapped twice, on July 21 and October 6, for 36 minutes total.
Fireworks AI listed 210 API incidents, Google's Gemini API page listed 2, and Together AI listed zero. Those numbers say more about posting habits than uptime.
49% of OpenAI and Anthropic API incidents started on a weekday between 9 AM and 5 PM Eastern, which is only 24% of the week.

How the Incidents Were Counted

The window is July 14, 2026 at 00:00 UTC through October 12 at 00:00 UTC, and every page was read on October 12. Ninety days is not arbitrary: OpenAI's page runs on incident.io, which keeps 90 days of history, and so do Groq, Cohere, Perplexity and Fireworks AI. Anthropic and Cerebras use Atlassian Statuspage, and Google posts Gemini API status on Google AI Studio. xAI's and Mistral's pages answered our scripts with Cloudflare bot checks, so they are missing.

An incident counts as an API incident only if the provider attached it to a model API component: OpenAI's APIs group without the dashboard Login component, Anthropic's Claude API (api.anthropic.com), and Google's API and Multimodal Live API components. Four incidents with no component were classified from their text. ChatGPT, claude.ai, Claude Code, Codex, billing and support incidents are in the totals but not in the API numbers.

Hours run from the provider's reported start to the update that restored the API component. A reopened incident uses its first resolution, and overlapping incidents from the same provider are merged so an hour is never counted twice. Error spikes nobody declared do not exist in this data.

OpenAI vs Anthropic vs Gemini

Provider	All incidents	API incidents	API hours	Median	Longest	Page's own uptime
OpenAI	109	20	72.7	124 min	20.6 h, Responses API latency, Aug 31	99.95%
Anthropic	82	60	76.5	51 min	6.0 h, multiple models, Aug 5	99.54%
Gemini API	4	2	48.0	2.8 h and 45.2 h	45.2 h, Batch API, Sep 14	bars, no number

OpenAI and Anthropic finished within 4 hours of each other, by opposite routes. Anthropic usually puts the affected model in the title ("Elevated errors on Claude Sonnet 5"), and it logged three times as many API incidents as OpenAI, each one shorter. OpenAI logs fewer and keeps them open: "Elevated latency in the Responses API" ran 20.6 hours starting the evening of August 31, and "Errors processing large PDFs in the Responses API" ran 12.8 hours on October 6.

Severity tilts the other way. Anthropic marked the Claude API as a major outage 3 times and a partial outage 23 times, including a 2.5-hour "Elevated errors across all models" on July 29. OpenAI's API components peaked at partial outage, 3 times, and 15 of its 20 API incidents were labeled degraded performance. Anthropic's volume also dropped month over month: 31 API incidents in the last 18 days of July, then 17, 8, and 4 in the first 11 days of October. OpenAI went 8, 3, 7, 2.

Do not compare the uptime percentages. On October 12, OpenAI showed 99.95% for its APIs group and Claude's page showed 99.54% for the Claude API, even though their incident hours were nearly identical. Different status-page software computes those numbers over different component lists. Compare timelines instead. If you want to see what a percentage means in minutes per month, the Uptime SLA Calculator does the conversion.

The Two Overlaps

On July 21, 2026, between 9:04 and 9:22 AM ET, an OpenAI image generation incident had spread to its Images API while Anthropic was reporting elevated errors on Claude Haiku 4.5. On October 6, between 8:24 and 8:43 AM ET, OpenAI's Responses API was choking on large PDFs while Anthropic reported elevated errors for Claude Opus 5.5.

That is the whole list. According to the providers' own pages, a plain text request had a working provider to fail over to for every minute of the 90 days.

Gemini's overlaps look bigger, 230 minutes with OpenAI and 146 with Anthropic, but each one involves either the 45.2-hour Batch API delay that started September 14 or a voice-cloning failure on October 1. Neither affects a real-time text call.

Incident Counts Mostly Measure Posting Policy

The smaller providers make the point clearly:

Provider	API incidents	API hours	What the count reflects
Fireworks AI	210	181.8	207 are automated single-model "Service Degradation" alerts, median 12 min
DeepSeek	29	13.5	short API alerts, median 14 min, titled in Chinese and English
OpenRouter	3	4.4	includes a 4-minute burst of 429s on its Anthropic and OpenAI routes
Cerebras	3	5.0	one 5.0-hour disruption on September 25
Perplexity	2	1.7	most of its 8 incidents cover its Computer product
Groq	1	6.1	one model, minimax-m2.7, October 10
Cohere	1	4.4	one multi-endpoint disruption, September 1
Together AI	0	0	nothing posted, though its monitors show 99.061% to 100% per model
Vertex AI	0	0	the public feed had 7 incidents total, none for Vertex AI

Fireworks AI posted 52 times as many incidents as the Gemini API page, which does not make it 52 times less reliable. Together AI posted nothing while its own Better Stack monitors show Qwen3.8-2.4T-A95B at 99.061% over 90 days. A raw count rewards whoever posts least. Google's page is also candid about priority: "Free-tier requests use sheddable capacity, while billed-tier requests are protected by critical priority."

Incidents Cluster in the US Workday

Out of 80 OpenAI and Anthropic API incidents, 39 started Monday through Friday between 9 AM and 5 PM Eastern. Those 40 hours are 24% of a 168-hour week but caught 49% of incidents. Part of that is just when engineers are around to declare something. Either way, that is when your fallback should be tested.

What It Means for Your Fallback

A retry loop with exponential backoff buys you seconds. It will not outlast a 51-minute median incident, let alone a 124-minute one. Run your own schedule through the Exponential Backoff Calculator and the gap is obvious. Past the retry budget, the request needs a different destination.

Fail over to another provider, not just another model. On July 29, Anthropic's outage hit every model for 2.5 hours, so falling back from Opus to Sonnet would have failed too, while OpenAI would have answered. Route by feature as well: every overlap involved images, PDFs, batch or voice on at least one side. Those workloads deserve their own fallback, their own health check, and a deadline on batch jobs so a 45-hour delay does not go unnoticed.

You can skip writing the failover yourself. Vercel AI Gateway tries a models array in providerOptions.gateway in order, OpenRouter has a models parameter that does the same, and LiteLLM takes a fallbacks list per model. The original post links a full comparison of LLM gateways.

Monitoring What the Status Page Leaves Out

Your own 500s may come from short spikes nobody ever declares. Upptime (MIT, 17,178 stars on October 12, 2026) runs checks from GitHub Actions every 5 minutes, which is the tightest schedule Actions allows, opens an issue when a check fails, closes it on recovery, and publishes a status page on GitHub Pages. A config for both APIs, with keys stored as repository secrets:

secrets:
  - OPENAI_API_KEY
  - ANTHROPIC_API_KEY

sites:
  - name: OpenAI API
    url: https://api.openai.com/v1/models
    headers:
      - "Authorization: Bearer $OPENAI_API_KEY"
    maxResponseTime: 5000
  - name: Claude API
    url: https://api.anthropic.com/v1/models
    headers:
      - "x-api-key: $ANTHROPIC_API_KEY"
      - "anthropic-version: 2023-06-01"
    maxResponseTime: 5000

Listing models costs no tokens, but it only proves the API responds. To catch the per-model failures Anthropic tends to post, switch an entry to method: POST and send a one-token request to the model you actually use.

Rerun the Numbers

Each CSV row is one incident: UTC times, API impact windows, severity, components, our classification note and a link to the original post. The original article includes a short Python script that merges each provider's windows and finds the overlaps. On the published file it prints:

OpenAI               72.7 h with an API incident open
Anthropic            76.5 h with an API incident open
Google Gemini API    48.0 h with an API incident open
OpenAI + Anthropic: 36 min at the same time
OpenAI + Google Gemini API: 230 min at the same time
Anthropic + Google Gemini API: 146 min at the same time

If you would classify an incident differently, open it from the url column, empty its api_windows_utc cell and run the script again.

What to Do With It

Single provider, no fallback: expect 73 to 77 hours per 90 days with an API incident open, roughly 3.5% of the time. Add a second provider before you spend time tuning retries.
Already on a gateway: pair OpenAI with Anthropic. Their 36 minutes of overlap means each covers the other for text.
Batch, PDF, image or voice: that is where both overlaps landed. Give each one a separate fallback and a deadline check.
Choosing a vendor: skip incident counts and uptime percentages, and compare hours on the components you call.

Wrapping Up

Three months of status pages put OpenAI and Anthropic about 4 hours apart, each with an API incident open around 3.5% of the time, and never broken on the same kind of request at once. Incident counts mostly reflect how a provider writes its status page. The fix for an LLM outage is a second provider behind a gateway plus a check you run yourself, not a longer retry loop. The full dataset and analysis are free to rerun and cite.

DEV Community