OpenAI's Rate Limit Reset Today Is a Signal, Not Just a Housekeeping Event
OpenAI reset Codex rate limits for all paid plans today — and the fact that a community member had to post about it publicly because the communication was unclear tells you most of what you need to know about how the agent infrastructure layer is being managed. The rate reset itself is not unusual operational housekeeping. What is notable is the timing: it comes less than a week after GPT-5.5 landed in the Codex runtime, and within 24 hours of an elevated-error incident that kept GPT-5.5 users seeing intermittent "not found" responses. OpenAI's status page confirms the incident resolved after about an hour on April 24. But the pattern — new model ships, errors appear, rate limits reset — is becoming a recognizable rhythm, and it is not a confidence-building one.
The token-based pricing that OpenAI rolled out across its plans makes this especially worth understanding precisely. Under the current rate card, GPT-5.5 costs 125 credits per million input tokens and 750 credits per million output tokens. For context, that output rate is six times the input rate — a meaningful signal about where the actual computation lives in a coding-agent context, where the model is generating far more tokens than it reads. By comparison, GPT-5.4 costs 62.50 credits input and 375 output. The delta is substantial, and it means that teams running heavy Codex workflows are burning through their plan allocations faster than the legacy per-message model implied. The limited-time offer currently running — 25x Plus limits through May 31 instead of the standard 20x — is OpenAI essentially buying developer goodwill while the pricing model settles. After May 31, that offer ends and the real cost structure takes over.
The practical implication for practitioners is not complicated: if you are running Codex in any kind of production workflow, assume rate limit variability is a permanent operational condition, not an exception to be handled reactively. The rate reset today is not a crisis. It is a data point. OpenAI is shipping new models into the Codex runtime on a cadence that is generating enough production load to require active limit management — which means the agent infrastructure itself is still hardening in real time, alongside the developers who depend on it. The community post that prompted this story was written by someone who found the communication inadequate and posted publicly because there was no clear reference. That is a reasonable reaction from a developer who just had their workflow interrupted and could not find a clear explanation. It is also the kind of friction that erodes trust faster than any benchmark differential.
The competitive frame is hard to ignore, even if it is not the story itself. Both major agent coding platforms — GitHub/Copilot and OpenAI/Codex — are simultaneously managing infrastructure instability while adjusting their economics. GitHub's post this morning described a 30x scaling problem driven by agentic workflow growth. Today's Codex rate reset is a smaller, more contained event, but it points in the same direction: the agent runtime layer is under production load that its management infrastructure is still catching up to. Teams that selected a single agent tool based on benchmark superiority are discovering that runtime predictability — not just raw capability — is the operational variable that matters most for production use. The model that wins on a leaderboard but runs into limit variability in your actual CI pipeline is not the better tool. That is a useful calibration for anyone making build-versus-buy decisions in the current market.
Sources: OpenAI Developer Community, OpenAI Status Page, OpenAI Help Center