OpenAI and Anthropic both publish uptime percentages on their status pages, and they publish them differently: OpenAI shows one figure per component for a 90-day window plus a figure for each group of components, while Anthropic shows one figure per component per month. We wanted to know whether those numbers can be checked against the incident records behind them, and whether they can be compared. They can be checked. Both sets of figures are closely reproduced by one rule: a full or major outage minute counts as downtime, each partial-outage minute counts as 0.3 of a downtime minute, and degraded performance counts as none.
Using that rule on the same window, 1 August to 10 October 2026, Claude’s uptime is lower than the closest ChatGPT component in each of four product pairs, but across each company’s components taken together the gap narrows to 0.17 points, and OpenAI’s chat and API products spent more time in a degraded state, which uptime ignores.
Key takeaways
- Both pages’ published uptime closely fits one rule: major or full outage 100%, partial outage 30%, degraded performance 0%. Rebuilt that way, Anthropic’s 15 component-months (August to October) differ from the published figure by 0.018 percentage points on average (at most 0.088; the two largest gaps, 0.088 and 0.074, are both in August), and OpenAI’s 17 components by 0.005 (at most 0.009). Counting partial outages in full misses by 0.63 and 0.10 points on average. In a sweep of weights from 0 to 1, 0.30 fits best for both, and 0.25 and 0.35 fit worse.
- The documentation of the platform OpenAI’s page runs on lists partial outage as “down” and mentions no weighting, but OpenAI’s published numbers behave as if partial outages count at 30%, which is what Atlassian’s documentation says for Anthropic’s platform. We cannot explain the difference.
- Uptime ignores degraded performance. Over 1 August to 10 October, ChatGPT Conversations spent 4.5% of the time in a degraded-only state and claude.ai 1.2%, and neither lowers its uptime.
- On the same rule and window: claude.ai 99.72% against ChatGPT Conversations 99.94%, Claude Code 99.64% against Codex 99.94% (99.83% with Codex in ChatGPT Desktop added), Claude’s API 99.74% against Chat Completions and Responses at 100.00%. For all 16 ChatGPT components together, 99.69% against claude.ai’s 99.72%. Which looks better depends on whether you count degraded time, which components you pick, and how each company chooses to mark statuses.
Last reviewed 10 October 2026. OpenAI’s figures were read from status.openai.com (its page data is timestamped 05:29 UTC) and Anthropic’s from status.claude.com/uptime, both on 10 October. The platform documentation was retrieved, and a copy archived, the same day. Our rebuilt figures come from each company’s incident records and are our own calculation, described in the method section, with the data linked.
What do the two pages publish?
OpenAI’s page shows a rolling period headed “Jul 2026 – Oct 2026”, which we read as a 90-day window to 10 October (a reading the fit below supports), with a figure for each component and each group: APIs (13 components) 99.96%, ChatGPT (16 components) 99.52%, Codex (4 components) 99.95% and FedRAMP 100%. Its footnote says availability is reported at an aggregate level across all tiers, models and error types. Among the components, Conversations shows 99.91%, ChatGPT Work 99.87%, Codex in ChatGPT Desktop 99.89% and Image Generation 99.79%.
| Anthropic component | August 2026 | September 2026 | October 2026 (to 10 Oct) |
|---|---|---|---|
| claude.ai | 99.59% | 99.69% | 99.92% |
| Claude Console (platform.claude.com) | 99.92% | 100% | 98.80% |
| Claude API (api.anthropic.com) | 99.72% | 99.73% | 99.92% |
| Claude Code | 99.59% | 99.69% | 99.92% |
| Claude Cowork | 99.56% | 99.70% | 99.92% |
| Claude for Government | 100% | 100% | 100% |
Anthropic’s page shows these values to two decimals without rounding up: claude.ai’s underlying August value, for example, is 99.59997%. The headline number also depends on what you pick. ChatGPT is 99.52% as a group but 99.91% for Conversations, and the reason is in the platform’s documentation: incident.io’s page says a group counts as up only when every component in it is up, so one component down takes the whole group down. OpenAI’s page footer says “Powered by” and links to incident.io, and its incident permalinks are on incident.io. When we rebuilt the ChatGPT group from its 16 components’ records on the rule below, we got 99.51%, close to the published 99.52%. The same group read 99.65% on 4 October, when we last checked it for our comparison of incident counts, and the fall of 0.13 points is of the size you would expect from the roughly 560 minutes of partial outage marked on 7 and 9 October, although the 90-day window also moved by six days.
Can the published numbers be reproduced?
The two platforms document different rules. Atlassian’s Statuspage documentation (retrieved 2026-10-10), which Anthropic’s page runs on, says only time in the major outage or partial outage state counts, that partial outages are “discounted to only be 30% as bad as major outages”, and that degraded performance is not considered. incident.io’s documentation (retrieved 2026-10-10) lists full outage and partial outage as “down” and degraded performance as “up”, and does not mention any weighting.
We tested both readings against the numbers. For each provider we rebuilt every component’s status history from its incident records: for Anthropic, the component status changes recorded in each incident’s updates, and for OpenAI, the status and time range each incident set on each component. Where several incidents overlapped on a component we took the worst status at each moment. Then we calculated uptime under two rules, partial outage at 30% and partial outage in full, and compared both with the published figures.
| Anthropic component and month | Published (underlying) | Rebuilt, partial at 30% | Rebuilt, partial in full |
|---|---|---|---|
| claude.ai, August | 99.600% | 99.674% | 99.100% |
| claude.ai, September | 99.700% | 99.700% | 98.999% |
| Claude API, August | 99.725% | 99.690% | 98.965% |
| Claude API, September | 99.734% | 99.734% | 99.114% |
| Claude Code, August | 99.599% | 99.511% | 98.557% |
| Claude Code, September | 99.700% | 99.700% | 98.999% |
| Claude Cowork, August | 99.567% | 99.511% | 98.557% |
| Claude Cowork, September | 99.706% | 99.694% | 98.979% |
| Claude Console, August | 99.927% | 99.927% | 99.927% |
| Claude Console, September | 100% | 100% | 100% |
| Claude Console, October (10 days) | 98.807% | 98.807% | 96.022% |
The remaining October values match just as closely: claude.ai, the API, Claude Code and Cowork are each 99.928% published and rebuilt, and Console’s 98.807% is reproduced exactly, a strong test because counting partial outages in full would give 96.02%. Across all 15 component-months from August to October, the rebuilt figures differ from the published ones by 0.018 percentage points on average and by at most 0.088 when partial outages count at 30%, against 0.63 on average and 2.79 at most when they count in full. The two largest gaps are in August, where the rebuild is above the published figure for claude.ai and below it for Claude Code, so missing downtime alone cannot explain them, and we have not identified the cause.
| OpenAI component or group (90 days to 10 Oct) | Published | Rebuilt, partial at 30% | Rebuilt, partial in full |
|---|---|---|---|
| Conversations | 99.91% | 99.913% | 99.709% |
| Login (ChatGPT group) | 99.97% | 99.972% | 99.908% |
| ChatGPT Work | 99.87% | 99.869% | 99.564% |
| Codex in ChatGPT Desktop | 99.89% | 99.892% | 99.638% |
| Image Generation | 99.79% | 99.788% | 99.294% |
| Voice mode | 99.96% | 99.962% | 99.875% |
| Connectors/Apps | 99.95% | 99.948% | 99.827% |
| Codex Web | 99.95% | 99.957% | 99.957% |
| Login (APIs group) | 99.995% | 99.995% | 99.985% |
| ChatGPT group (16 components) | 99.52% | 99.507% | 98.358% |
| APIs group (13 components) | 99.96% | 99.959% | 99.864% |
| Codex group (4 components) | 99.95% | 99.957% | 99.957% |
For OpenAI’s 17 components (the nine shown above, Images, and the seven that share 99.98%, with the two components named Login kept apart), the 30% rule is within 0.01 points of the published figure for every one, 0.005 on average, while counting partial outages in full is within 0.02 points for only 9 of the 17 and misses by up to 0.50. Seven of the components (Compliance API, Search, File uploads, GPTs, Deep Research, Agent and Sites) share identical counted outage time, 46.4 minutes of partial outage, so the real number of separate tests is nearer eleven. All seven land 0.009 above the published 99.98% and Codex Web 0.007 above its 99.95%, while four other components land 0.001 to 0.002 below theirs, so we cannot tell whether OpenAI cuts or rounds. The three groups are within 0.013, though they are not independent of the components: the APIs group equals Images and the Codex group equals Codex Web. So whatever the platform’s documentation says, the numbers on OpenAI’s page behave as if partial outages count at 30%. We do not know why: it may be a setting on OpenAI’s page, or the documentation may be out of date, and we have seen neither confirmed.
| Partial-outage weight | Anthropic: mean difference (largest), points | OpenAI: mean difference (largest), points |
|---|---|---|
| 0 (ignored) | 0.261 (1.193) | 0.050 (0.210) |
| 0.2 | 0.082 (0.398) | 0.020 (0.069) |
| 0.25 | 0.040 (0.199) | 0.012 (0.034) |
| 0.30 | 0.018 (0.088) | 0.005 (0.009) |
| 0.35 | 0.057 (0.199) | 0.010 (0.037) |
| 0.5 | 0.187 (0.796) | 0.029 (0.143) |
| 1.0 (counted in full) | 0.634 (2.785) | 0.103 (0.496) |
The sweep shows 0.30 as the best fit for both providers, and neighbouring weights fit visibly worse, so the rule is well supported by these two pages. It does not tell us the weight is exactly 0.30, or why OpenAI’s page behaves this way.
What does uptime leave out?
Degraded performance counts for nothing under this rule, and it is a large share of the time. Over 1 August to 10 October, ChatGPT Conversations was in a degraded-only state for about 4,520 minutes, 4.5% of the 70 days, with 186 minutes at partial outage. claude.ai was degraded-only for about 1,190 minutes (1.2%), with 833 minutes at partial outage and 36 at major outage. Claude Console, similarly, was degraded for about 240 minutes in September yet shows 100% uptime for the month.
That is why we show the chart above. The bars show the share of time each product spent in any marked problem state, including degraded. On that measure OpenAI’s chat product (4.67%) and API products (2.71%) look worse than Claude’s (2.04% and 2.19%), while Claude’s coding tool looks worse than Codex’s four components (2.19% against 0.73%) but not once Codex in ChatGPT Desktop is added (2.67%), and Claude’s agent product (8.01%) looks worse than ChatGPT Work (4.13%). Claude Cowork’s figure comes mostly from a single Windows-update incident that was marked degraded for about 99 hours, which we covered in our article on it.
How do the two compare on the same rule and window?
| Product pair | Anthropic: uptime | Anthropic: degraded-only time | OpenAI: uptime | OpenAI: degraded-only time |
|---|---|---|---|---|
| claude.ai vs ChatGPT Conversations | 99.72% | 1.18% | 99.94% | 4.49% |
| claude.ai vs all 16 ChatGPT components together | 99.72% | 1.18% | 99.69% | 10.42% |
| Claude Code vs Codex (4 components) | 99.64% | 1.09% | 99.94% | 0.68% |
| Claude Code vs Codex plus Codex in ChatGPT Desktop | 99.64% | 1.09% | 99.83% | 2.23% |
| Claude API vs OpenAI Chat Completions and Responses | 99.74% | 1.32% | 100.00% | 2.71% |
| Claude API vs all 13 OpenAI API components together | 99.74% | 1.32% | 99.99% | 3.48% |
| Claude Cowork vs ChatGPT Work | 99.64% | 6.90% | 99.88% | 3.72% |
| Any of Claude’s 5 components vs any of 33 OpenAI components | 99.46% | 10.38% | 99.63% | 12.43% |
All figures are for 1 August to 9 October 2026 inclusive (70 days), which we chose to line up with Anthropic’s calendar months, and they use the rule above. Pairing products is our judgement: claude.ai is one component while ChatGPT’s chat experience is spread across many, so we show both Conversations alone and all 16 together, and the Codex and API pairs each with and without the extra components. Claude has no equivalent grouping, so the last row shows each company’s components taken together, with the 33 OpenAI components being the 16 ChatGPT, 13 API and 4 Codex ones. We would not read any single row as a verdict.
To put the percentages in minutes, 99.9% uptime over 30 days is about 43 minutes of counted downtime, 99.7% is about 130 minutes, and 99.5% is about 216 minutes. Our earlier comparison of incident counts reached a similar conclusion about counts: they depend heavily on how each company chooses to report.
What this means if you rely on these tools
- Treat a published uptime figure as a statement about outage time only. Degraded performance, which can be several percent of the time, is excluded on both pages.
- Check which component and window a number describes. OpenAI’s headline ChatGPT figure (99.52%) is for 16 components together and a 90-day window, while Anthropic’s figures are monthly and per component.
- Compare like with like, or not at all. On our same-window rebuild neither provider is ahead everywhere, and the ranking changes with whether degraded time counts and which components are included.
- Remember who sets the statuses. Each company’s staff decide when to mark a component partial or major outage or merely degraded, as we saw in our check of OpenAI’s status page against its own write-ups, so the figures reflect what each company chose to record.
What we could not verify
- We cannot say what users experienced. Uptime here means time in a status each company’s staff set, not measured availability.
- We do not know why OpenAI’s published figures behave as if partial outages count at 30% when the documentation of the platform its page runs on lists them as down with no weighting. We observed the match, not its cause.
- Our Anthropic rebuild uses the component status changes recorded in incident updates, so changes made outside incidents would not appear. Where several incidents overlap on a component we take the worst status, whereas Statuspage keeps one current status per component, so a component’s status may revert earlier on the real page than in our rebuild.
- We assumed OpenAI’s window is the 90 days ending when we read the page (after 05:29 UTC on 10 October), and Anthropic’s October figure covers 10 days. Both pages’ figures move every day, so you will see different numbers if you check now.
- Anthropic’s feed covers only its 50 most recent incidents (from 27 July) and OpenAI’s covers incidents resolved since 10 July, so our same-window comparison starts on 1 August, which we chose to match Anthropic’s calendar months. The 10 October pull of OpenAI’s feed starts on 12 July, because earlier incidents had rolled off.
- Two intervals, one per provider, belong to incidents with no resolved update when we read the pages, and we held them open to the read time. One is Claude Console at degraded performance from 7 October, about 3,000 minutes inside our same-window span, which accounts for about 3 points of the 10.38% degraded-only figure in the last row (it would be about 7.4% without it). The other, a delayed-data notice on OpenAI’s Compliance API, affects no figure here.
- The product pairings are our judgement and the two companies split their products into components differently. Seven OpenAI components share identical counted outage time, so our OpenAI test is nearer eleven independent checks than seventeen. We did not rebuild Claude for Government, whose three figures are all 100%.
- Anthropic’s figures are cut to two decimals (we saved the raw values) and OpenAI’s rounding is unclear, so differences under about 0.01 points are not meaningful.
How we researched this, and the data
On 10 October 2026 we read OpenAI’s status page for its published uptime and Anthropic’s per-component uptime pages, saving the underlying monthly values, and archived both platforms’ uptime documentation. We took Anthropic’s component status changes from the updates in its incident feed and OpenAI’s component statuses and time ranges from each incident’s page, keeping only the impacts that belong to that incident, since every OpenAI incident page also embeds the status page’s currently open incident, and keeping the two components both named Login apart. Scripts (dataset_openai.py and uptime_rebuild.py) rebuilt each component’s status timeline, took the worst status where incidents overlapped, calculated uptime under both rules and a sweep of weights, and compared the results with the published figures. The 499 status intervals we exported, 335 from OpenAI and 164 from Anthropic, are in this CSV, and OpenAI’s per-incident data as of 8 October (111 incidents, without per-component statuses) is in this one. The interval CSV is the one that lets you recompute the uptime figures. For the wider picture of OpenAI’s incidents see our dataset of 111 incidents, and for what the “resolved” stage means, our study of status-page stages.
Frequently asked questions
OpenAI’s status page showed 99.52% for the ChatGPT group of 16 components over what appears to be the 90 days to 10 October 2026, and 99.91% for Conversations alone. Both exclude degraded performance, and partial outages appear to count at 30%.
Anthropic’s status page showed claude.ai at 99.59% for August 2026, 99.69% for September and 99.92% for October to 10 October, and the Claude API at 99.72%, 99.73% and 99.92%. These are monthly, per component, and cut rather than rounded to two decimals.
On Atlassian Statuspage, which Anthropic uses, major outage minutes count in full, partial outage minutes at 30% and degraded performance not at all. OpenAI’s page, which runs on incident.io, documents partial outages as down with no weighting, but its published figures match a 30% weighting in our rebuild. In both cases degraded performance does not reduce uptime.
We cannot say. Rebuilt on the same rule for 1 August to 10 October 2026, Claude’s uptime is lower than the closest ChatGPT component in each of four product pairs, but across all of each company’s components the gap narrows to 0.17 points (99.46% against 99.63%), OpenAI’s chat and API products spent more time in a degraded state, and the answer changes with which components you compare and whether degraded time counts.
