AI API Reliability Report: OpenAI, Claude Status History

Status pages answer one question well: is it broken right now. They answer the next one badly: how often does it break. Ask the Claude status API for its incidents and you get the latest 50, which on 4 October 2026 reached back only to 25 July. OpenAI's returns 25, which reached back to 14 September. Ask either about last spring and the record is gone from the API.

So I started keeping it. Once a day a script reads 12 public status pages, stores each incident once under the provider and the incident's own id, and never deletes one. Where a page still exposes older history, I pulled that in too. This page is what the archive says, rebuilt every day.

Gravity's agents run on model APIs like these, which is why their incidents are the first thing I read in the morning. The numbers below are the providers' own words, counted carefully, and nothing more.

The last 90 days

Model APIs first, then the agent and automation platforms that sit one layer up. For OpenAI, Anthropic, Perplexity and Vercel only the API part of the status page counts; the consumer apps on the same pages are left out. The coverage table further down says exactly which components count for each service.

The 90 days to 4 October 2026. Incidents are counted by start date. Hours are the time at least one in-scope incident was open, clipped to the window. "Data from" marks a service whose usable history starts inside the window, so its figures cover less than 90 days. What counts for each service is listed under coverage. Every incident is in data.csv.

The last 30 days

The same table over a shorter window, which is the one to quote when a writer needs "recently".

The 30 days to 4 October 2026, same rules as the 90-day table.

Leaderboard: this month and last

Ranked by the fewest hours with an open incident, so the top of each table is the quietest status page. Early in a month these tables are thin, and a single incident can move a service several places, so the full previous month sits beneath for comparison.

Model APIs, October 2026 so far

Ranked by fewest hours with an open incident, then fewest incidents; ties share a rank. October 2026 from the 1st to 4 October 2026.

Agent and automation platforms, October 2026 so far

Ranked by fewest hours with an open incident, then fewest incidents; ties share a rank. October 2026 from the 1st to 4 October 2026.

Model APIs, September 2026

Ranked by fewest hours with an open incident, then fewest incidents; ties share a rank. The full month of September 2026.

Agent and automation platforms, September 2026

Ranked by fewest hours with an open incident, then fewest incidents; ties share a rank. The full month of September 2026.

Month by month

The last seven months for every service. This is the part no status page shows you, because their APIs forget and their history pages are one provider at a time.

Each cell: incidents started that month, and below it hours with an open incident. Tracking since 4 October 2026, when the daily collector first ran; earlier months come from each provider's own retained history: back to 1 October 2025 for the Atlassian Statuspage pages (Claude, Make, Cursor, Vercel), about 90 days for the incident.io pages (OpenAI, Zapier, Perplexity, Cohere, Groq, Fireworks AI, v0; Cohere, Groq, Zapier and v0 reach further through their latest-25 list), and whatever Google's incidents feed still lists. "From" marks a month the history only partly covers; n/a means no usable history. Full series since 1 October 2025: monthly.csv.

How to read these numbers

Hours with an open incident are not downtime. They measure how long the provider's own status page showed at least one in-scope incident as unresolved. Most incidents are partial: slower responses, errors on one model, one region or one feature. A status page also tends to open an incident after the trouble starts and to close it when the provider is confident, so the window runs long in both directions. Treat the figure as a reporting measure.

Providers post at different thresholds. Fireworks AI posts an incident per affected model, so it logs far more incidents than a provider that groups problems into one notice. A page that posts every slowdown will look worse here than one that posts only big outages. That is a real difference in transparency, and it cuts the opposite way from how a raw ranking reads.

Scope matters more than it looks. The OpenAI and Claude rows count only incidents that touched an API component, not ChatGPT or claude.ai. The Zapier, Make, Cursor and v0 rows count every incident on the page, because each page describes one product. Zapier and Make also post incidents for single third-party connectors (one app's trigger failing, say), which can stay open for days and push their hours up even when the core product is fine. If you quote a figure, quote the scope with it.

Zero means nothing was posted. It is not a test result. If you run production traffic on any of these, the practical move is a fallback chain rather than a bet on one provider: I wrote up how to set up agent fallback chains and retry and fallback patterns, and the agent uptime and reliability guide covers what to measure on your own side. Rate limits fail differently from outages, which the rate-limiting post covers.

Change log: notable incidents

Every major or critical incident, plus anything open six hours or more, in the last 90 days, newest first. Each links to the provider's own incident page.

1. 1 October 2026: Claude API (Anthropic), minor impact: Delayed credits on the Claude Platform, open 6 h 16 min.
2. 29 September 2026: Cursor, major impact: Investigating service degradation - xAI models, open 35 min.
3. 29 September 2026: v0 by Vercel, major impact: Upstream Provider issue causing partial outage in v0, open 34 min.
4. 29 September 2026: Claude API (Anthropic), major impact: Elevated errors on claude.ai, Claude Code, Claude Cowork and the Claude API, open 2 h 6 min.
5. 22 September 2026: Claude API (Anthropic), major impact: Elevated errors for multiple models, open 1 h 38 min.
6. 18 September 2026: Cursor, major impact: Investigating service degradation, Grok Bot, open 5 h 43 min.
7. 18 September 2026: Make, major impact: Degradation on us1.make.celonis.com, open 3 h 56 min.
8. 18 September 2026: Cursor, major impact: Investigating service degradation, Grok Bot, open 2 h 56 min.
9. 17 September 2026: Cursor, major impact: Investigating service degradation, Grok Bot, open 2 h 4 min.
10. 16 September 2026: Cursor, major impact: Investigating service degradation, Grok Bot, open 6 h 30 min.
11. 15 September 2026: OpenAI API, major impact: Elevated errors with gpt-image-2.5-flare, open 40 min.
12. 15 September 2026: Claude API (Anthropic), major impact: Intermittent error spikes for Claude Mythos 5.1 and Claude Fable 5.1, open 1 h 26 min.
13. 11 September 2026: Claude API (Anthropic), major impact: Elevated errors for Claude Mythos 5.1 and Claude Fable 5.1, open 17 min.
14. 9 September 2026: Zapier, major impact: Zap History not displaying Zap runs, open 1 h 50 min.
15. 9 September 2026: Make, major impact: Degradation on EU1 Region, open 18 h 17 min.
16. 8 September 2026: OpenAI API, minor impact: Elevated errors for image generation, open 7 h 27 min.
17. 7 September 2026: Cursor, major impact: Grok Bot computers unreachable for some users, open 1 h 5 min.
18. 3 September 2026: Cursor, major impact: Degraded performance for Grok Bot routine triggers, open 1 h.
19. 3 September 2026: Cursor, major impact: Elevated errors for Anthropic Models, open 2 h 15 min.
20. 3 September 2026: Cursor, critical impact: Investigating service degradation, open 3 h 26 min.
21. 3 September 2026: Claude API (Anthropic), major impact: Elevated errors for multiple models, open 2 h 57 min.
22. 3 September 2026: Claude API (Anthropic), major impact: Elevated errors for Claude Sonnet 5, open 19 min.
23. 2 September 2026: Claude API (Anthropic), major impact: Elevated errors for Claude Sonnet 5, open 26 min.
24. 2 September 2026: Cursor, major impact: Investigating service degradation, Automations, Cloud Agents, Review Agents, open 1 h 20 min.
25. 1 September 2026: Zapier, no impact level posted: Freshservice - Webhook delivery disruption, open 6 d.
26. 31 August 2026: OpenAI API, minor impact: Elevated latency in the Responses API, open 20 h 38 min.
27. 27 August 2026: Cursor, critical impact: Investigating service degradation, cursor.com and IDE, open 7 min.
28. 24 August 2026: Claude API (Anthropic), major impact: Elevated errors for multiple models, open 3 h 24 min.
29. 21 August 2026: Zapier, no impact level posted: Salesforce integration - trigger and connection disruptions, open 4 d 4 h.
30. 20 August 2026: Claude API (Anthropic), major impact: Elevated errors on requests to multiple models, open 26 min.
31. 4 October 2026: Tracker. Tracking started. Backfilled Claude, Make, Cursor and Vercel from their status-page history to 1 October 2025; incident.io pages (OpenAI, Zapier, Cohere, Groq, Fireworks AI, Perplexity, v0) from their 90-day windows and latest-25 lists; Google Cloud from its incidents feed.

Notable means major or critical impact, or open for 6 hours or more, in the 90 days to 4 October 2026; newest first, up to 30 lines. The page rebuilds daily, so this list moves.

Coverage and status at the last check

What counts for each service, how far back the usable history goes, how many incidents the archive holds, and what each status page said when the collector last read it.

Status as read at 13:36 UTC on 4 October 2026. This is a once-a-day snapshot, not a live monitor; for "is it down right now", open the status page linked in the row.

Methodology

Here is exactly how the tables are built, so you can decide how far to trust them.

Where the data comes from

Atlassian Statuspage pages (Claude at status.claude.com, Make, Cursor, Vercel): the public /api/v2/incidents.json, /api/v2/components.json and /api/v2/summary.json endpoints every day. For the backfill and to close any gap, the page's own /history.json listing and the per-incident /incidents/<id>.json record, which carries start, resolve and component data. These pages were backfilled to 1 October 2025.

incident.io pages (OpenAI, Zapier, Cohere, Groq, Fireworks AI, Perplexity, and v0 at v0-status.com): the Statuspage-compatible /api/v2/incidents.json and /api/v2/summary.json, plus the JSON the status page itself loads, which lists every incident in its 90-day window with the affected components and impact times. The compatible API carries no component data, so components come from the second source.

Google Cloud: the public incidents feed at status.cloud.google.com/incidents.json, filtered to incidents that list a Vertex AI or Gemini product. Google's severity labels are mapped to this page's scale as service information to minor, service disruption to major, service outage to critical. That mapping is mine.

What counts

An incident is anything the provider posted as an incident. Scheduled maintenance is excluded. Each incident is stored once, keyed by status page and the provider's incident id, and later reads update it in place, for example when it resolves. Impact is the provider's own label (none, minor, major, critical) where the page gives one; for older incident.io records outside the compatible API's latest 25, it is derived from the worst component status posted (degraded performance to minor, partial outage to major, full outage to critical), a mapping that matched the provider's own label on 81 of the 83 incidents where I could compare both on 4 October 2026.

How each number is computed

Incidents are counted in the window where they started. Major or critical is the subset with that impact label. Hours with an open incident is the union of each incident's start-to-resolve interval, clipped to the window, so two overlapping incidents count once. Median time to resolve uses incidents that started in the window and have resolved. Longest incident is the longest start-to-resolve span among incidents that started in the window, linked to the provider's page. Start is the provider's stated start time where the page records one (Statuspage's started_at, the earliest component impact on incident.io), otherwise the time the incident was posted. Windows end at the time of the last collection.

Some incidents are posted after the fact, and a page that moves to a new status provider can import its old incidents with the import time as the post time. Either way the record shows a resolve time earlier than its start. Those incidents still count, dated at their resolve time, but they add no hours and no time to resolve, because the real start is unknown. The CSV flags each one.

The collector runs daily. Nothing is ever deleted from the archive, which lives in the site repository with one dated snapshot per day, so a figure quoted from this page can be checked against the day it was taken.

Known gaps

I checked 19 status pages on 4 October 2026 and kept only those with a working public JSON feed. Not included: n8n Cloud and Lindy (no Statuspage JSON at their status pages), Relevance AI and Together AI (no JSON feed found), Replit and Mistral (requests refused), OpenRouter (status JSON unavailable when checked), and the Gemini API in Google AI Studio, whose status page renders in the browser with no public JSON feed I could find. I will add any of them when a feed works.

incident.io pages expose only their last 90 days with component data, so OpenAI and Perplexity history here starts in early July 2026 and grows from there. Google's feed listed six incidents across all of Google Cloud on 4 October 2026, the oldest from 27 February 2026, so Gemini on Vertex AI has a short record. None of this measures performance from the outside: there are no synthetic pings here, only what each provider published.

FAQ

Sources