Codex users report outages, 5-hour caps, and token shortages after Sites launch

• OpenAI shipped Codex's June 2 announcement with Sites, annotations, and six role-specific plugins, then spent the next day dealing with reliability complaints that thsottiaux's incident post later summarized as three separate small incidents.
• Users spent June 2 and June 3 reporting a mix of outright failures and quota pain: doodlestein's outage report showed cross-machine errors, while bridgemindai's earlier limits complaint and doodlestein's six-account post described a sudden step change in Codex availability.
• OpenAI staff also confirmed the 429 spike in public: according to the GitHub incident thread, maintainers said it was an active incident and not a client-side issue, which lines up with jxnlco's too-many-requests note.
• The quota complaints were specific, not generic grumbling. bridgemindai's follow-up said a 5-hour limit was exhausted in about an hour after the Codex 2X promo ended, and doodlestein's Claude Code comparison spelled out how parallel agents and subagents can turn rate limits into a hard stop.
• OpenAI responded by resetting usage limits across paid plans in thsottiaux's reset announcement, and users like deredleritt3r's /goal reply immediately reported much longer runs.

You can read OpenAI's Sites and plugin launch post, the public 429 incident thread on GitHub, and a fresh OpenAI community post about unexpected quota resets. The launch itself added six role-specific plugins across 62 apps and 110 skills in OpenAI's telling, while the rough edges that surfaced right after included repeated 429s, disappearing token headroom, flaky tool connections, and even a PowerShell input corruption bug.

OpenAI's launch was not a small UI tweak. In the official announcement, the company said Codex now has more than 5 million weekly users, that non-developers are about 20% of that base, and that Sites is rolling out in preview for Business and Enterprise workspaces.

The June 2 ship had three concrete parts:

• Sites: interactive hosted websites and apps, shareable by workspace URL.
• Annotations: point-and-edit refinement for sites, docs, spreadsheets, and slides.
• Six role-specific plugins: data analytics, creative production, sales, product design, public equity investing, and investment banking.

OpenAI also said those plugins bundle 62 apps and 110 skills, and named early Sites ecosystem partners including Vercel, Wix, Replit, Lovable, Figma, and Webflow in the same post. The timing matters because the reliability blowback landed almost immediately after this wider, less developer-only rollout.

The cleanest evidence of the outage wave is that users, OpenAI's GitHub maintainers, and OpenAI leadership all described the same thing in public.

doodlestein's June 2 post said the errors were hitting all of his machines and different accounts. Hours later, jxnlco's short update said OpenAI was aware of another Codex issue involving too many requests.

Then, in the public GitHub issue, OpenAI maintainers told users there was an active incident, that it was not a client-side issue, and later that the incident had been mitigated. By early June 4, thsottiaux's statement put the count at three separate small incidents over the prior 24 hours.

That sequence turns a pile of "is anyone else seeing this" posts into a pretty direct incident timeline: cross-machine failures on June 2, 429 acknowledgments on June 3, then a platform-wide reset after leadership said the service had taken three hits in a day.

The louder complaint was not a single outage. It was that Codex suddenly felt stingier.

The reports clustered around three symptoms:

• 5-hour bucket pain: bridgemindai's follow-up said the 5-hour limit was burned in about an hour after the Codex 2X promo ended.
• Weekly quota anxiety: bridgemindai's earlier post said a ChatGPT Pro account had already consumed 45% of a weekly limit after two days on GPT 5.5.
• Account-hopping as a workaround: doodlestein's six-account post said one dev machine went through six GPT Pro 20x accounts in a day, which he described as a step change from the prior week.

OpenAI has an official Codex rate card, but the useful detail in the evidence pool is not the pricing table. It is that users were watching both a short 5-hour bucket and a weekly bucket, then seeing behavior they thought had changed materially from the previous week. A new OpenAI community thread about an unexpected reset shows at least one user noticing that a quota page displaying roughly 70% remaining later reset to a new cycle after the incident response.

OpenAI's response was blunt. In thsottiaux's reset announcement, the company said it had reset usage limits for Codex across all paid plans. Replies like dkundel's tokens-flow reply and deredleritt3r's /goal reply suggest the extra headroom was visible right away.

The other theme in user reports was not raw uptime. It was workflow friction around how Codex decides, asks, and disconnects.

Two patterns stand out:

• Tool instability: koltregaskes's Linear post said the integration would disconnect mid-session, then fail to reconnect before working again minutes later.
• Too much second-guessing: the same post complained that Codex kept asking for approvals on actions needed to finish its own task, while bclavie's workaround showed a more psychological version of the same problem, where Codex resisted an optimization until the prompt reframed the bad code as something written by an intern.

That connects back to koltregaskes on Codex as the foundation, which argued that Codex's agent capabilities are the part worth preserving if OpenAI keeps collapsing Codex and ChatGPT together. The complaint was less about chat quality than about whether the agent loop stays in flow.

One useful comparison came from outside OpenAI. doodlestein's Claude Code post described Anthropic's own rate-limit failure mode in detail, and the parallels are awkwardly close.

His post separated two different bottlenecks that often get mashed together in casual discussion:

• Usage limits: how much budget you can burn over a 5-hour or weekly window.
• Rate limits: how many requests you can fire in a shorter interval.

That distinction matters because Anthropic had just spent June 1 fixing a bug where, according to ClaudeDevs' reset note, some Claude Code sessions spawned excessive parallel subagents and burned usage faster than expected. In a follow-up, ClaudeDevs' explanation said the problem was Opus 4.8 request handling that triggered more parallel tool calls than intended.

Codex users were describing a different product, but a very similar shape of pain: agentic coding tools get much more fragile when parallelism, background tools, and quota accounting stop lining up.

One more concrete bug surfaced away from the outage noise. AlemTuzlak's PowerShell report said recent Codex releases made Windows PowerShell terminals unusable by corrupting terminal input.

That report is only one post, so it does not establish a broad Windows regression on its own. It does add a separate category of breakage, though: beyond backend incidents and token scarcity, at least one user was reporting a local terminal-level failure tied to recent client releases. For a week dominated by quotas and 429s, that is a different problem class entirely.