Quota went up. Miles per gallon went up too.
In late August 2026, OpenAI Codex lead Tibo Sottiaux did two things in one announcement cycle: reset paid ChatGPT Work and Codex weekly usage limits for subscribers, and ship a batch of harness fixes — compaction, memory, goals, automations, subagents, computer history, rolling summaries, and MCP — that OpenAI claims deliver 10–50% efficiency gains on the same quota.
That pairing matters. A limit reset without harness improvements is a gift that expires on the next long loop. Harness fixes without a reset are invisible until you hit the wall again. Together they are OpenAI's answer to a month where Anthropic is cutting net Claude Code capacity and Cursor users are planning around an OpenAI shutoff — keep Codex subscribers whole and make the harness do more per token.
TL;DR — what changed
| Question | Answer |
|---|---|
| What did Tibo reset? | Paid ChatGPT Work + Codex weekly usage meters |
| What shipped besides quota? | Fixes across compaction, memory, goals, automations, subagents, computer history, rolling summaries, MCP |
| Claimed efficiency gain? | 10–50% more effective work on the same quota (OpenAI/Tibo figure) |
| Is this a new plan tier? | No — manual top-up + product fixes, not a SKU change |
| Relation to 15M/20M milestones? | Same reset pattern, new fix batch — see 15M and 20M banked reset posts |
| 5-hour window still active? | Yes — weekly and 5-hour meters both bind unless your UI says otherwise |
| Best spend? | Long agent loops on work that was quota-blocked |
| Cursor users affected? | Indirectly — OpenAI models in Cursor still face Nov 12 shutoff; Codex CLI is the durable OpenAI path |
The fix batch — what each piece actually does
Tibo's list is not a feature dump — it is the agent harness operational stack, the 98%+ of Codex that is not the model call itself:
| Component | Why it affects quota efficiency |
|---|---|
| Compaction | Summarizes context without losing task state — fewer restarts after long sessions |
| Memory | Recalls project facts across threads — less re-explaining in every loop |
| Goals | Keeps multi-step tasks aligned — fewer wasted tool calls on drift |
| Automations | Repeating workflows without re-prompting — lower setup tax per run |
| Subagents | Parallel workstreams with isolated context — faster completion without one mega-thread |
| Computer history | Continuity across shell/GUI actions — less re-discovery of environment state |
| Rolling summaries | Incremental context compression — delays hard /clear moments |
| MCP | Tool server reliability and discovery — fewer failed or redundant tool calls |
OpenAI's 10–50% range is plausible if you were losing loops to compaction bugs, amnesiac memory, or MCP timeouts. It will be invisible on "write me an email" prompts. It shows up on the overnight Codex loops explainx.ai tracks — migrations, test suites, multi-package refactors.
The reset — how it differs from prior gifts
Tibo has reset paid limits before — notably the August 9 reset after the Claude Code proxy ban incident, and the August 13 15M-user refill. Pattern:
- Manual top-ups, not a permanent raise to the weekly formula
- Often paired with a milestone or product moment
- 5-hour and weekly meters both still apply after the gift
The August 21 20M banked reset shifted toward user-scheduled refills. This late-August batch appears to be another immediate meter refill plus the harness fix ship — check your Work and Codex usage panels to confirm it landed on your account.
Not announced: pay-to-reset SKU changes (OpenAI is testing paid resets separately), permanent weekly formula increases, or an Astra model drop tied to this post.
Tibo vs. the "5% of Cursor traffic" debate
The same week, Cursor CEO Michael Truell said OpenAI models are only ~5% of Cursor traffic — a figure Tibo pushed back on indirectly, arguing tokens are not a proxy for revenue or value. That exchange matters for Codex users separately from Cursor:
- High-value OpenAI workflows (long Codex agent runs, GPT-5.x on hard tasks) can be a small share of requests but a large share of tokens and margin.
- Codex CLI and ChatGPT Work are where OpenAI keeps the direct relationship — no Cursor in the path, no November shutoff.
- If you were on OpenAI models inside Cursor, this reset does not save that integration — plan migration per the Cursor shutoff guide.
What to run now — copy-paste priorities
1. Confirm both meters after the reset:
codex /status
(or check ChatGPT Work usage in the web UI — Codex and Work share the agentic pool on paid plans).
2. Re-run blocked loops. Anything that died on "weekly limit reached" mid-migration — resume with /continue if the session checkpoint exists.
3. Use Fast tier on supported models. Tibo's prior resets ended with go /fast — toggle Fast in the CLI when the catalog advertises a Fast tier for your model.
4. Test subagents on parallelizable work. Security review + test generation + docs update in one goal, split across subagents — that is where the fix batch should show up in wall-clock time.
5. Compare against Claude Code post-September 14. If your team runs both harnesses, re-baseline after Anthropic's limit swap — Codex may be the higher-capacity side of the pair for the rest of Q3.
Honest limitations
- 10–50% is a vendor claim until independent before/after benchmarks exist — task-dependent, hard to audit from outside.
- Resets are gifts, not policy — the July restored 5-hour window can tighten again if demand spikes.
- Cursor OpenAI access still ends November 12 regardless of Codex CLI generosity.
- Performative reset criticism from the community (resets landing right after weekly renewal) still applies — verify the meter actually moved before celebrating.
Related on explainx.ai
- Codex 15M users — Tibo's reset tease
- Codex 20M users — banked reset
- Tibo vs Boris — harness freedom and the August 9 reset
- How to run loops in Codex CLI
- Codex as a platform — OpenAI's open agent harness
- OpenAI ends Cursor model access — November 12 migration
- Claude Code limits drop 17% September 14 — the opposite move
- Claude Code vs. Codex rate limits
- Loop engineering with coding agents
Primary sources:
- Tibo Sottiaux (@thsottiaux), late August 2026 posts on Codex limit reset and harness fix batch (X).
- Michael Truell (@mntruell), August 29, 2026 — ~5% OpenAI traffic in Cursor (context for same-week debate).
Reset timing and fix details are accurate as of August 30, 2026. Confirm your usage meters in ChatGPT Work or Codex CLI before planning long loops. OpenAI may ship additional fixes without a separate announcement.