AI API outages on September 3, 2026
UTC day · measured every 5 minutes from 4 regions · method
On September 3, 2026 (UTC) our probes attributed 4 outage(s) to 2 provider(s): Gemini API (Google): 3; SambaNova API: 1. Providers’ own status pages opened 2 entries (Anthropic).
This page is history. Is it down right now?
Live status, updated every 30 minutes: Gemini API (Google), SambaNova API, Claude API (Anthropic), friendli API · all AI APIs, last 72 hours
Was the Gemini API (Google) down on September 3, 2026?
Yes, measured. 3 outage(s) attributed to Google: requests to gemini-flash-lite-latest failed from up to 3 of 4 regions at once while no other provider was failing, first at 14:25 UTC, at least 42 minutes observed in total. At least once Google’s own host answered with an HTTP 5xx.
| UTC | What failed | Regions | Duration | Failed probes | Verdict | Error |
|---|---|---|---|---|---|---|
| 08:00–08:13 | requests to gemini-flash-lite-latest | 3 of 4: Asia (Tokyo), Europe (Germany), US (Central) | ≥13 min | 3 | shared with 1 other provider(s) — not attributed | The read operation timed out |
| 13:10–13:35 | requests to gemini-flash-lite-latest | 2 of 4: South America (São Paulo), US (Central) | ≥25 min | 3 | shared with 1 other provider(s) — not attributed | The read operation timed out |
| 14:25–14:40 | requests to gemini-flash-lite-latest | 3 of 4: Asia (Tokyo), Europe (Germany), South America (São Paulo) | ≥15 min | 3 | provider outage | The read operation timed out; http 503 |
| 15:15–15:55 | requests to gemini-flash-lite-latest | 4 of 4: Asia (Tokyo), Europe (Germany), South America (São Paulo), US (Central) | ≥40 min | 6 | shared with 1 other provider(s) — not attributed | The read operation timed out; timeout |
| 16:15–16:35 | requests to gemini-flash-lite-latest | 3 of 4: Europe (Germany), South America (São Paulo), US (Central) | ≥21 min | 5 | provider outage | timeout |
| 17:10–17:15 | requests to gemini-flash-lite-latest | 2 of 4: Asia (Tokyo), Europe (Germany) | ≥6 min | 2 | provider outage | timeout |
Is the Gemini API (Google) down right now? Live status →
Was the SambaNova API down on September 3, 2026?
Yes, measured. 1 outage(s) attributed to SambaNova: host probes failed from up to 2 of 4 regions at once while no other provider was failing, first at 00:48 UTC, at least 2 minutes observed in total. SambaNova’s status page lists nothing that day.
| UTC | What failed | Regions | Duration | Failed probes | Verdict | Error |
|---|---|---|---|---|---|---|
| 00:48–00:50 | host probes | 2 of 4: Europe (Germany), South America (São Paulo) | ≥2 min | 2 | provider outage | timeout |
Is the SambaNova API down right now? Live status →
Was the Claude API (Anthropic) down on September 3, 2026?
Not by our measurements. No failure of Anthropic was seen from two or more regions at once. Anthropic’s own status page opened 2 entries that day (self-reported, listed below).
Self-reported on Anthropic’s status page, not verified by our probes:
- major Elevated errors for Claude Sonnet 5 · resolved Sep 3, 12:56 UTC
- major Elevated errors for multiple models · resolved Sep 3, 16:23 UTC
Is the Claude API (Anthropic) down right now? Live status →
Was the friendli API down on September 3, 2026?
Not attributably. 1 multi-region failure(s) of friendli coincided with failures of other providers, so the cause sat above friendli and is not blamed on it.
| UTC | What failed | Regions | Duration | Failed probes | Verdict | Error |
|---|---|---|---|---|---|---|
| 17:55–18:05 | host probes | 4 of 4: Asia (Tokyo), Europe (Germany), South America (São Paulo), US (Central) | ≥11 min | 5 | shared with 2 other provider(s) — not attributed | timeout |
Is the friendli API down right now? Live status →
51 single-region failure(s) were also recorded that day (yi-01ai 18, google 16, sambanova 6, nebius 3, iflytek 2, sarvam 2, baichuan 1, deepinfra 1, writer 1, qwen 1). A failure seen from one region cannot be separated from a problem on that network path, so none of them is called an outage.
How is an outage told apart from a network problem?
Each provider is probed every five minutes from 4 independent regions. A failure is attributed to the provider only when it is seen from two or more regions at once and no other provider fails in the same minutes; a failure shared with other providers has a cause above all of them and is not blamed on anyone. Durations are lower bounds: we see the first and last failed probe, not the real start and end. Status-page entries are the provider’s own words, shown next to the measurement, not instead of it.
← Sep 2 · Last 72 hours · Sep 4 →
Machine-readable: incidents.json · Licence: CC BY 4.0