Across the 90 days ending September 29, 2026, OpenAI recorded 99.34% uptime and one major or critical incident, while Anthropic, Perplexity and Cohere each recorded 100.0% uptime and 0 min of downtime. OpenAI's single incident lasted 56 minutes and was the one incident marked critical out of the 10 incidents StatusBird logged across all monitored services in the last 30 days.
StatusBird checks each of these services every 2 minutes from its own probes, and every figure below comes from those readings over the 90-day window ending September 29, 2026. The full definition of a check and a grade is on the methodology page.
What do the 90-day numbers show?
| Service | Uptime | Grade | Major/critical incidents | Total downtime | Average incident | Last incident |
|---|---|---|---|---|---|---|
| OpenAI | 99.34% | B | 1 | 56 min | 56 min | 2026-09-25 |
| Anthropic | 100.0% | A+ | 0 | 0 min | 0 min | none in window |
| Perplexity | 100.0% | A+ | 0 | 0 min | 0 min | none in window |
| Cohere | 100.0% | A+ | 0 | 0 min | 0 min | none in window |
Every number in that table favours Anthropic, Perplexity and Cohere. Three of the four providers went the whole window without a major or critical incident. OpenAI is the one with a mark against it, and its grade of B is the second lowest grade in StatusBird's entire 90-day table, behind the C grades held by Plaid at 98.73% and Shippo at 98.93%.
Why is OpenAI at 99.34% if it only lost 56 minutes?
Because the two columns measure different things. The incident count and the downtime total cover readings serious enough to be classified major or critical. The uptime percentage counts every reading that was not fully operational, including partial degradation that never escalated into a named incident. A provider can therefore show a short incident list and still carry a lower uptime figure, which is exactly the shape of OpenAI's 90 days: one 56 minute incident, plus a run of degraded readings spread across the window.
That distinction matters for how you plan. A provider that fails hard and rarely is easy to fail over from, because your error handling has a clean signal to react to. A provider that returns slow or partial responses without going fully dark is harder, because your code often gets a technically valid answer that arrives too late for a shopper waiting on a chat reply. Anthropic, Perplexity and Cohere gave StatusBird neither kind of reading in this window.
What did the OpenAI outage on September 25 look like?
It started at 22:59 UTC on September 25, 2026 and recovered at 23:55 UTC the same evening, for 56 minutes at a worst status of critical. Two other services had incidents within hours on either side of it: ShipStation lost 254 min earlier that day, from 07:57 UTC to 12:11 UTC, and Plaid lost 94 minutes on September 26 from 03:41 UTC to 05:16 UTC. If your store runs an AI support agent and a shipping integration and a bank-linked payment method, that 48 hour stretch touched three separate parts of your stack.
The timing is worth noting for anyone running AI-backed customer chat. A 22:59 UTC start is late evening in Europe and mid to late afternoon in North America, so the incident straddled a live selling window on one side of the Atlantic and an unattended overnight window on the other. Current readings for each provider sit on their live pages: OpenAI, Anthropic and Perplexity.
Which stores does the difference actually matter for?
Stores with AI in the checkout or support path
If an LLM call sits between a shopper and an answer, the gap between 100.0% and 99.34% is a real gap. A chat widget that hangs on a degraded API looks broken to the customer even though your storefront is fine. For these stores, the provider choice is a customer experience decision, and the 0 min downtime recorded for Anthropic, Perplexity and Cohere is a meaningful advantage over OpenAI's 56 minutes plus degraded readings.
Stores using AI for batch or back-office work
If you generate product descriptions, translate a catalogue, classify support tickets overnight or summarise reviews on a schedule, a 56 minute outage costs you a retry, not an order. A job that fails at 22:59 UTC and runs again after 23:55 UTC lands before most merchandising teams open their laptops. For this pattern the difference between the four providers is close to irrelevant, and you should pick on cost, model quality and rate limits instead.
Agencies running AI features for multiple clients
One provider incident hits every client at once. With OpenAI at 1 major/critical incident and the other three at 0, an agency standardising a single provider across a portfolio is making a concentration bet. Splitting clients across providers does not reduce total incidents, but it does mean a 56 minute event affects part of your book rather than all of it.
How does AI uptime compare with the rest of your stack?
Context helps here. Over the same 90 days, Shopify recorded 99.9% with 124 min of downtime across 2 incidents, ShipStation recorded 99.8% with a single 254 min incident, and Plaid recorded 98.73% with 1642 min across 5 incidents. OpenAI's 56 minutes of major and critical downtime is small next to any of those. The B grade reflects the degraded readings, not a catastrophic outage record.
None of the four AI providers appear on StatusBird's current above-normal risk list. That list is led by Vercel at a modelled 51.8% chance of a non-operational reading in the next 24 hours, followed by TikTok Ads at 37.2% and BigCommerce at 34.0%. Monthly totals for the wider stack are in the outage index, where August 2026 recorded 16 major/critical incidents and 40.4 hours of combined downtime.
What should you change if AI is in your customer path?
- Put a hard timeout on the call. Decide how long a shopper waits before your widget gives up, and set it in the client that makes the request rather than relying on the provider default.
- Configure a second provider in the same code path. On timeout or a 5xx response, retry once against a second key. With three providers sitting at 100.0% and 0 min of downtime in this window, there is no shortage of a second option.
- Write the fallback copy now. Replace the AI reply with a specific line that offers the contact form or live agent queue, not a spinner. Store it as a constant so nobody has to edit code mid-incident.
- Move anything non-urgent into a queue. Product description generation, ticket tagging and review summaries should retry on a schedule rather than fail in the moment. A 56 minute incident then costs one retry cycle.
- Cache the answers you repeat. Size guidance, shipping policy and returns answers rarely change. Serving them from cache means a provider incident does not take out your most common questions.
- Get told when it happens. Watch the provider you actually use, with alerts routed to the person who can flip the fallback switch.
If you want an independent reading on any of these services rather than waiting for a vendor status page to catch up, StatusBird checks them every 2 minutes and can send alerts to SMS, email, Slack, Teams or Discord. You can set that up here.
Frequently asked questions
Which AI provider had the best uptime over the last 90 days?
Anthropic, Perplexity and Cohere tied at 100.0% uptime with 0 major or critical incidents and 0 min of downtime. OpenAI recorded 99.34% and grade B.
How long was OpenAI's outage in September 2026?
It lasted 56 minutes, starting September 25, 2026 at 22:59 UTC and recovering at 23:55 UTC, with a worst status of critical.
Is a 99.34% uptime figure bad for an API my store depends on?
It depends on where the API sits. OpenAI's 56 minutes of major and critical downtime is far below Plaid's 1642 min or ShipStation's 254 min over the same window, but the 99.34% figure also counts degraded readings, which matter more when a shopper is waiting on a live response.
Are any AI services currently flagged as elevated risk?
No. The current above-normal list is led by Vercel at 51.8% and TikTok Ads at 37.2%, and none of the four AI providers appear on it.
How many incidents did StatusBird record across all services recently?
There were 10 incidents in the last 30 days and 37 major or critical incidents across the full 90-day window.