Skip to content

Does a 99.9% uptime promise mean your AI agent will finish?

Unity Horizon6 min read

No. One published SLA starts counting downtime above a 10% error rate held five minutes, and a 20-step agent run at 5% finishes 36% of the time.

No. Google's SLA for Gemini for Google Cloud does not begin counting downtime until more than 10% of requests fail, and only once that has held for five consecutive minutes. Everything under that line is uptime. A twenty-step agent run losing 5% of its calls finishes 36% of the time, inside a month the provider can report as perfect.

What does an AI provider count as downtime?

Google publishes two of them, and they put the line in different places.

Gemini for Google Cloud Vertex AI Platform
Stated objective 99.9% monthly uptime 99.9% for training, deployment and batch prediction, 99.5% for custom model online prediction on two or more nodes
Downtime begins error rate above 10% error rate above 5%
Shortest countable outage 5 consecutive minutes 1 minute
Responses that count 5XX carrying code "Internal Error" HTTP 500 or 503
Volume floor 2,000 valid requests in the measurement period none stated
Credit between 99.0% and 99.9% 10% of that month's bill 10% of that month's bill

The Gemini for Google Cloud SLA defines Downtime as "more than 10 percent Error Rate", in Downtime Periods of "five or more consecutive minutes". The Vertex AI Platform SLA says 5% and one minute. Neither document names a Gemini API model endpoint: the Vertex covered-service table lists training, deployment, batch prediction, AutoML and custom model online prediction, and the word Gemini appears nowhere in the body of it. Which of the two covers a generateContent call is a question for an account manager, and the time to ask is before you need the answer. Checked by Unity Horizon on 13 November 2026.

Does every provider publish an uptime number?

No. Anthropic's commercial terms, effective 17 June 2025, provide the services "as is" and "as available", and disclaim any warranty "that their use will be uninterrupted". No percentage appears in the document.

Search for one anyway. On 13 November 2026 the pages that rank for the question offered 99.5%, 99.9%, 99.95% and 99.99%, each stated flatly, none linking to a document that says so. A procurement questionnaire asking a supplier for its uptime figure will be answered with one of those four, copied out of a blog, and the answer will go in a file.

How often does an AI API log an incident?

The count below covers the Claude API alone, because its feed is the only one of the three we could count this way. We fetched status.claude.com/api/v2/incidents.json on 13 November 2026 and took every incident opened between 15 August and 12 November 2026, a ninety-day window the feed covers completely. The same list is readable at status.claude.com/history. OpenAI's feed, pulled the same morning, attaches no component to any of the 25 incidents it carries, and Google Cloud's returned eight incidents grouped by product, the earliest dated 27 February 2026. The three feeds record different things, so no total across them means anything.

Forty-five incidents. Twenty-seven of them listed the Claude API component, and every one of those was resolved: a median of 40 minutes, the shortest 6, the longest 376. Seventeen named a specific model in the title. Eighteen were filed as elevated errors and five as degraded performance.

Add every minute of all twenty-seven together and you get 2,047 minutes, 1.6% of the window. That figure is not an error rate and it is not downtime. It is the span during which the vendor said something was wrong, and seventeen times out of twenty-seven the something was one model.

Why does a 5% error rate break a twenty-step run?

Multiplication. An agent that calls a model twenty times needs twenty consecutive successes, and the per-call rate compounds against it.

Per-call failure rate 5 calls 10 calls 20 calls 40 calls
1% 95% 90% 82% 67%
5% 77% 60% 36% 13%
10% 59% 35% 12% 1.5%

At 5%, which neither definition above counts as downtime, 36 runs in 100 reach the end. At 10%, inside the Gemini for Google Cloud threshold, 12 do, and the provider closes the month at 100%. We published the same arithmetic from the model's side in How many times in a row does an AI agent get the same task right?, where a benchmark score of 37.4% on one attempt fell to 20.6% across four. Uptime is counted in minutes. A run is counted once.

Do retries fix it?

Partly, and least when it matters. The official Claude SDK retries transient failures with exponential backoff, twice by default, honouring the retry-after header. Most isolated failures never reach your code at all.

Retrying works on independent failures. A provider incident is the case where they are not independent: the same minute is bad for everyone calling that model, and three attempts inside it are three failures and a longer wait. The other half of the problem, what a retry costs when the first attempt already landed a write, is in How do you stop an AI agent from doing the same thing twice?.

Status codes decide the contract as well as the code path. The Claude API documents 429 rate_limit_error, and 529 overloaded_error for when "the API is temporarily overloaded". OpenAI documents 503 service_unavailable_error with code server_is_overloaded when the requested model is overloaded, and a 429 with code slow_down when your request rate climbs too fast. Apply either definition of Error Rate quoted above to that set and most of it drops out: a 429 is not a 5XX, and a 529 is a 5XX whose code is not Internal Error.

Who pays for the difference?

The thresholds were drawn for one shape of workload, and an agent is the other shape. A classifier, a reply draft, a label on a support ticket, a summary under a search box: one call, one person waiting, and a 5% error rate surfaces as a page somebody reloads. A filing is a chain. We built OFFSET, which files declarations directly to the ΑΑΔΕ portal, and a declaration is a sequence of form steps rather than a single request, which is the shape described in How do you automate a government portal that has no API?. The same error rate reaches the two of them as different numbers.

Volume sorts the same two groups again. The Gemini for Google Cloud SLA measures its error rate "subject to a minimum of 2000 Valid Requests in the measurement period". Send fewer than that and the minute your agent died is in nobody's denominator.

What should you measure instead?

Two numbers, both from your own client.

Start with the per-call error rate, split by model and by status code. The split is what makes it worth collecting: seventeen of the twenty-seven incidents in our window named one model, so a fallback to a second model at the same vendor clears most of them, and a second vendor clears the remainder. A fallback chosen after the fact clears nothing.

Then count finished runs over started runs, by day. That ratio is what a user experiences, no service level agreement describes it, and one week of it is worth more than a year of status-page archaeology.

We have not run a full quarter of client-side error logging against all three providers, so every count here is the vendor's own account of its own downtime, which is a weak form of evidence and the best that is public. If you hold client-side numbers for the same ninety days, we would like to see them.

If you are scoping an agent that has to complete a chain of steps, our agentic AI systems page sets out how we scope one. What share of your agent runs finished last week, and who counted them?

Common questions

What does 99.9% uptime mean for an AI API?

It is a promise about the provider's error rate per minute, not about your success rate. Under the Gemini for Google Cloud SLA, downtime starts above a 10% error rate sustained for five consecutive minutes, so a month in which 9% of requests fail in every minute is a month at 100% uptime.

Why does an AI agent fail when the provider's status page is green?

Because an agent run needs every one of its steps to succeed. At a 5% per-call error rate, which neither published definition of downtime counts, a 20-step run finishes 36% of the time. The status page counts minutes the provider called bad. Your run counts one failure.

Does Anthropic publish an uptime percentage for the Claude API?

Not in its public commercial terms, effective 17 June 2025, which provide the services 'as is' and 'as available' with no warranty that use will be uninterrupted. Its status page publishes incidents instead: 27 of the 45 logged between 15 August and 12 November 2026 listed the Claude API component, with a median duration of 40 minutes.

How should you measure AI API reliability?

From your own client rather than the vendor's page. Track the per-call error rate split by model and by status code, and track the ratio of completed agent runs to started runs. Those two numbers describe your traffic, and no service level agreement describes either.