ChatGPT Work and Codex — OpenAI's paid agent surfaces for multi-step tasks and CLI coding — hit elevated errors and failed sessions across September 25–26, 2026. Agent runs stalled mid-tool-call, Work mode tasks failed to start, and status dashboards showed a partial outage rather than a full platform blackout.
OpenAI Codex lead Tibo Sottiaux responded the way heavy Codex users have learned to expect in 2026: acknowledge the pain, name a concrete fix timeline, and promise a usage limit reset for paid Codex and Work subscribers caught in the window. That reset promise lands in awkward context: on September 24, OpenAI paused paid weekly reset purchases — the optional "refill my quota now" checkout that had been testing at $5–80 depending on tier per explainx.ai's pay-to-reset coverage.
If you orchestrate production work through Codex, this week is a stacked lesson: hosted agents fail, goodwill resets are not a SLA, and the paid escape hatch can disappear the day before an outage.
TL;DR
| Question | Answer |
|---|---|
| What broke? | ChatGPT Work + Codex — elevated errors, failed agent tasks (Sept 25–26, 2026). |
| Who promised relief? | Tibo Sottiaux — automatic usage reset for paid Codex/Work users. |
| Paid reset purchases? | Paused Sept 24 — buy-to-refill checkout unavailable before outage. |
| Same as Aug 31 Work outage? | No — August incident was Work-centric; this hit Codex/agent paths harder. |
| Fallback? | Pre-tested second harness; watch Claude Code limits if switching. |
| Pattern? | Post-incident Tibo reset — see Astra postmortem reset. |
Timeline — September 24 through 26
Understanding the sequence prevents mixing three separate policy events into one narrative:
| Date (2026) | Event | Why builders care |
|---|---|---|
| Sept 24 | OpenAI pauses paid weekly reset purchases (checkout disabled for quota top-ups) | Removes the flat-fee "refill now" option documented in August pay-to-reset testing |
| Sept 25 (evening UTC) | Status: ChatGPT Work / Codex elevated errors | Agent tasks fail even when chat still loads |
| Sept 26 | Mitigation rolls out; partial recovery; Tibo posts reset promise | Goodwill quota credit — not a capacity upgrade |
| Ongoing | Users watch whether reset stacks with banked resets | OpenAI historically leaves scheduling questions unanswered |
The September 24 pause is easy to miss because it is not an outage headline. It is a pricing/control change: OpenAI stepping back from selling instant refills at the same time subscriber frustration about Astra-era limit cuts was still fresh.
Then the platform hiccupped — and the usual monetary relief valve was already closed.
What users saw during the outage
Reports clustered into four buckets consistent with prior Work-mode incidents:
- Work tasks fail to enqueue — spinner, generic error, no partial output saved.
- Codex CLI auth succeeds but tool execution loops error — especially long-horizon runs.
- Intermittent success by region — typical of partial outages, not a clean global off switch.
- Standard ChatGPT chat still usable for many accounts — reinforcing that the agent layer is a separate reliability domain.
That last point matters for architecture reviews. If your "AI feature" is really a Codex or Work backend, your uptime is not ChatGPT's headline uptime.
Tibo's reset promise — pattern, not guarantee
Tibo Sottiaux has become the public face of post-incident quota make-goods for Codex users — a role explainx.ai documented through the 15M user reset tease, Astra banked reset, and September 12 quality postmortem reset.
The September 26 promise fits the same template:
- Acknowledge agent surfaces broke for paying users.
- Commit to an automatic usage reset (wording implied account-side credit, not a manual coupon code).
- Leave scheduling details fuzzy.
What it is not:
- A permanent raise in weekly caps.
- A refund for wall-clock downtime.
- A substitute for multi-harness redundancy.
Treat Tibo resets like airline vouchers after a cancellation: appreciated, not a reason to skip travel insurance.
September 24 — paid reset pause in plain language
Since August 2026, some ChatGPT paid tiers saw a pay-to-reset button after exhausting weekly agent quota — flat dollars to jump back to 100% instead of waiting for the calendar reset. explainx.ai mapped tested price bands up to $80 on the $200 Pro plan in OpenAI starts selling usage resets.
On September 24, 2026, user-facing checkout for those purchases paused — in-app copy described a temporary disable while OpenAI adjusted billing, not the end of the feature forever. Exact corporate wording varied by client build.
Operational impact:
- Heavy users who had incorporated buy-a-reset into launch-week planning lost that lever before the outage.
- Free banked resets (save-and-spend perks from launch promos) remained a separate mechanic — see banked reset explainer for how confusing the two meters already was for subscribers.
- Outage goodwill resets from Tibo are yet a third category — automatic credits tied to incidents.
If your finance team models AI spend, export three line items: subscription, optional reset purchases, incident credits. Collapsing them hides volatility.
Reliability context — GPT-6 Astra launch month
September 2026 was never going to be a quiet quota month. explainx.ai already covered:
- Astra launch and staggered access
- Full banked reset when rollout finished early
- Heavy-user limit cuts ~48 hours later
- Sol/Luna same-day launches and banked-reset meme wars
An outage plus reset promise at month-end reads less like random bad luck and more like capacity stress on agent tiers still absorbing Astra demand — even if OpenAI's eventual postmortem cites a narrower root cause (database, feature flag, regional routing).
Unrelated but emotionally adjacent: rumors of extra GPT-6 Pro / Max tiers in the UI do not fix Codex CLI downtime. Model-picker gossip and harness uptime are different on-call rotations.
What to do when Codex or Work fails mid-run
Before the next outage
- Keep a second harness warm — credentials tested, repo cloned, one successful
/commit-scale task completed this week. - Snapshot agent state — branch name, last prompt, partial diff — before starting irreversible migrations.
- Read both vendors' limit calendars — switching to Claude without checking September 14 Claude Code cap math can mean trading downtime for a weekly ceiling surprise.
- Do not depend on pay-to-reset — it was pausable on 24 hours' notice.
During the outage
- Check OpenAI status for component scope — Work-only vs Codex vs global.
- Pause scheduled loop runs that assume Codex tool access.
- Fail over to backup harness for read-only tasks first (review, planning) before write tasks (merges, deploys).
- Watch Tibo's account for reset confirmation — apply goodwill credits before launching another multi-day burn.
After recovery
- Verify whether an automatic reset landed — usage meters in Codex settings, not just chat.
- File internal incident notes: wall-clock delay, not only token spend.
- Re-evaluate whether agentic work belongs on hosted-only paths without a queue/retry layer you control.
Claude Code limits — fallback is not free
Teams that flipped to Claude Code during the September outage hit a different quota story. Anthropic replaced a temporary 50% weekly boost with a permanent 25% raise on September 14 — which is still a ~17% net cut versus promo-era capacity for many Max users. explainx.ai walked through the arithmetic in Claude Code limits drop 17%.
Anthropic did ship a UX fix the same week: sessions no longer hard-stop mid-response at the 5-hour window — see mid-task cutoff fix. That makes Claude a gentler fallback once you are running, but it does not expand weekly headroom.
Practical rule: maintain two tested harnesses whose combined weekly caps cover your critical path — not "Codex primary, Claude panic button."
What people are asking
Will OpenAI extend the reset to free tiers? Tibo's wording targeted paid Codex and Work users — consistent with prior incident credits. Free tiers historically do not receive the same quota refunds.
Does pausing paid resets mean OpenAI is ending overages entirely? Unknown. Pause ≠ permanent removal. It may be billing-system work, policy review, or reaction to backlash that pay-to-reset felt like double charging after summer's free banked resets.
Should enterprises escalate support tickets? For Business/Enterprise seats, yes — document failed Work tasks with timestamps. Goodwill resets may still flow through Tibo's public promise, but contract customers often need paper trails.
Is Codex less reliable than Claude Code? Wrong question. Both hosted harnesses fail. Compare your measured error rates and fallback readiness, not vibe charts from launch week.
Related reading on explainx.ai
- ChatGPT Work outage — August 31, 2026 timeline
- OpenAI $80 usage reset on the $200 Pro plan (pay-to-refill context)
- GPT-6 Astra quality postmortem — Tibo's Sept 12 reset
- Astra banked reset when rollout finished early
- Claude Code limits — September 14 net 17% cut math
- Claude Code mid-task cutoff fix at 5-hour reset
- GPT-6 Pro sightings — separate from Codex uptime
- Claude vs Codex vs Cursor usage limits — July reset wars
- Loop engineering for coding agents
- Why AI agents have not gone mainstream — reliability angle
Outage timestamps, reset promises, and pay-to-reset pause details reflect OpenAI status reporting and public posts by Tibo Sottiaux as of September 26, 2026. Billing mechanics may change without a dedicated changelog entry — verify in-app before planning spend.