OpenAI outages API and ChatGPT with elevated latencies across GPT-5.5 workflows

• OpenAI said it was "investigating elevated latencies across the board for API and ChatGPT," according to thsottiaux's incident post, then later marked the incident resolved in thsottiaux's resolution update.
• Community reports tied the disruption to GPT-5.5 workflows: Teknium's update said the issue did not look fully resolved early on, while SurafelDem's reply said GPT-5.5 was unusable through OpenAI even as other providers still worked.
• OpenAI narrowed the blast radius mid-incident when thsottiaux's scope clarification said the problem hit the completions API and did not affect Codex.
• That clean boundary did not fully match developer experience, because thdxr's OpenCode note reported stalled Codex subscription endpoints during the past week, and Teknium's Hermes update said OpenAI changed its spec to fix Codex OAuth breakage.

The official breadcrumbs were thin: the status page plus thsottiaux's first post confirmed elevated latency, and thsottiaux's follow-up closed it out hours later. Meanwhile the useful detail came from developers patching around broken streaming, including altryne's Hermes workaround, OpenCode reporting stalling in the field via thdxr's post, and Teknium noting that GPT-5.5 requests were still wobbling after the first round of fixes in his update.

OpenAI's public timeline was short. thsottiaux's incident post went up at 15:04 UTC saying API and ChatGPT were seeing elevated latencies, with the status page as the canonical update channel.

By 23:06 UTC, thsottiaux's resolution update reduced the public record to one word: "Resolved." That left most of the operational texture to user reports rather than the incident log.

The most concrete scope note came from thsottiaux's scope clarification, which said the outage was limited to the completions API and did not affect Codex.

Developer reports around GPT-5.5 were messier. In SurafelDem's reply, one user said GPT-5.5 was unusable through OpenAI while other providers still worked, and Teknium's follow-up answered bluntly that the issue sat with OpenAI rather than the user's own setup.

Even with the official scope narrowed, coding-agent tooling still showed adjacent breakage. thdxr's OpenCode note said Codex subscription endpoints had been unstable for a week and that OpenCode shipped patches in v1.15.11 to handle the stalling.

The sharpest symptom report came from Hermes. altryne's Hermes workaround described OpenAI Codex streams that delivered response.output_item.done events but sometimes ended without response.output, which in turn crashed openai-python 2.24.0 parsing and triggered model fallback. A few hours later, Teknium's Hermes update said OpenAI had fixed the Codex OAuth issue in the background and changed the spec enough that users needed to run hermes update.