How to Read the OpenAI Status Page's System Status and Uptime Figures

The OpenAI Status page (status.openai.com) answers two different questions, and reading it correctly means keeping them separate. The top banner tells you whether OpenAI currently knows about any problems ("We're fully operational" means no known issues affecting its systems). The uptime percentages below tell you how reliable each component has been over a selected date range. A component can show 100% uptime for the period and still be down right now, or show a dip from last month while working normally today. Check the banner first for "is it down right now," then the component list for "how reliable has this been."

What the overall status message means

The banner is a statement about known issues, not a guarantee. "We're fully operational" is paired with the wording "We're not aware of any issues affecting our systems" — so if you're seeing failures while the banner is green, the page simply hasn't registered the problem yet. In that case, the useful next steps are the component list (to see whether your specific product is flagged) and the history view, rather than assuming the page is wrong.

When an incident is active, the banner changes and the affected components are marked, so the banner and the component list should be read together.

The component groups

Components are organized into product families, each with its own component count and aggregate uptime:

Two practical points follow from this structure:

• Find your actual dependency, not the brand name. If you build on the API, "ChatGPT 99.70%" is not your number — look at the specific API component you call, such as Chat Completions or Embeddings. Likewise, a Codex CLI user should watch the CLI component rather than the Codex group average.
• Group figures are aggregates. The group-level percentage rolls up its components, so a single weak component can pull the group number down while everything you use is fine.

Reading the uptime percentages

Each component shows a percentage over the displayed range (the page shows a Jul 2026 – Oct 2026 window in the captured data). The figures are reported at an aggregate level across all tiers, models, and error types.

That last sentence is the most important caveat on the page, and the page states it directly: individual customer availability may vary depending on subscription tier and the specific model and API features in use. So a 99.97% API figure is a population-level statistic, not a promise about your account. Treat small differences between components — say 99.998% versus 99.99% — as noise rather than a meaningful reliability ranking, and treat the percentages as a way to spot which areas have had incidents, not to predict your own experience.

A rough way to translate the numbers: 99.99% over a month is a few minutes of disruption; 99.9% is closer to tens of minutes; 99.7% (the ChatGPT group figure in the sample) is on the order of a couple of hours. Use this only to gauge scale — the page does not break downtime into incident durations per component.

A worked example

Suppose your application calls the Embeddings API and your users report failures.

1. Check the banner. If it says fully operational, no incident is acknowledged yet.
2. Check the Embeddings component specifically. If it shows 100% for the period, the historical figure won't explain today's failure — the issue may be new, or on your side (credentials, rate limits, network).
3. Check adjacent components you depend on: Login (99.98% in the sample) if your failures are authentication-related, or Files if you upload before embedding.
4. Open the history view to see whether similar components had recent incidents that match your symptoms.

This sequence avoids the common mistake of reading a high uptime percentage as "everything is fine right now" or a group-level dip as "my specific endpoint is broken."

Common pitfalls

• Confusing "no known issues" with "no issues." The banner reflects OpenAI's awareness, which lags real-world failures.
• Using group uptime as your SLA input. The page explicitly says availability varies by tier, model, and feature; the aggregate is not a contractual figure for your use.
• Ignoring the date range. Percentages are computed over the selected window. A 100% reading for a quiet period says nothing about a busy one.
• Over-reading tiny gaps. Differences at the third decimal place are not a reliable signal of which component to trust more.

If you need to track changes over time, the page offers a subscribe-to-updates option and a history view; those are the right tools for following incidents rather than re-checking the same percentages.