OpenAI's Codex Has Shipped 10 Resets in 20 Days

OpenAI's Codex team is 21 days into a very public experiment: ship one real improvement every day for 28 days, or give users a full usage reset instead. The pledge, from Codex lead Tibo Sottiaux, was meant to show off engineering velocity. Instead, the numbers it's producing say something else — when Codex's own paying users get to vote, they want the reset over the feature almost every time.

What changed

On October 5, Sottiaux, who runs Codex and ChatGPT Work at OpenAI, posted a pledge on X: for the next 28 days, each day the team would either ship one improvement "relevant for most codex/work users" or issue a full usage reset.

Through day 20, the tally was almost even: 10 reset days against 10 improvement-only days, logged publicly by community trackers and a day-by-day pledge tracker on GitHub.

github.com/ikaijua/Awesome-AITools/discussions/1232

The improvements have been real but mostly incremental: Codex's default models ran about 50% faster on day one, context length jumped from 272k to 500k tokens on day seven, persistent prompt caching landed on day 16, and day 19 brought faster Dots sync plus a batch of CLI fixes. The reset days have been exactly that — OpenAI clearing paid users' weekly usage caps, with no feature attached.

On October 26, day 21, Sottiaux posted again, tipping the tally to 11 resets against 10 improvements:

"Day 21/ Wonder what happens the remaining seven days. But today is a reset."

Seven days remain before the pledge ends on November 1.

Why it matters

This isn't just a changelog. Each day's ship gets an informal approval score from the community tracker codex-resets.com, built off reaction counts on Sottiaux's posts, and the pattern is blunt. Reset days have landed between 82% and 91% approval. Feature-only days have scored as low as 8% (an in-app message-editing update on day 14) and 20-29% (the day 19 CLI and caching bundle). Only one standalone feature, day 12's cache and context-reuse work, cleared 80%.

That gap is the actual story here. OpenAI set out to prove Codex ships fast. What the pledge has mostly proven is that the thing its paying users want most isn't a new feature — it's more room to run the one they already have. For a product whose usage limits were tight enough that OpenAI felt it needed a month-long public pledge to manage the backlash, that's an awkward thing to measure in the open.

It's also happening inside a specific competitive window. The pledge started weeks after a round of Codex Pro-plan quota cuts, and in the middle of Anthropic's Claude Code gaining ground with developers. One developer-tooling blog, Tech Insider, reported Claude Code ahead of Codex in a usage survey this year, 65.5% to 29.5%; treat the exact figure as one outlet's read rather than an industry-wide count, but the direction matches what shows up across the trackers covering this pledge.

tech-insider.org/ca/claude-code-vs-codex-2026/

What people are saying

Sottiaux's own framing, from the original pledge:

"Over the next 28 days, each day we'll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin."

Not everyone's convinced the pledge is delivering what it promised. Developer Priy (@alwayspriyesh) summed up the day-19 mood:

"Tibo's 28 days of Codex improvements feels like PR stunt... Atp, the reset button deserves employee of the month."

Finding developers publicly cheering the pledge is harder than finding skeptics — most of the enthusiasm shows up as approval votes on reset days rather than named posts, and OpenAI hasn't commented on why resets keep outscoring shipped features. That's a thinner sourcing picture than the skeptic side, and it's worth saying plainly rather than papering over.

What it looks like in practice

If you're running Codex inside an automated pipeline — a CI job, or an n8n workflow that calls the Codex CLI to open pull requests overnight — the pledge means you can't assume a stable weekly quota through November 1. Teams scheduling nightly codegen runs are left with two real options: batch the expensive jobs for right after a reset lands (recent resets have posted in the morning Pacific time, per Sottiaux's own timestamps), or build retry-with-backoff into the workflow so a mid-run quota error doesn't fail the whole job.

Community trackers like codex-resets.com and the GitHub pledge log have, unofficially, become part of the operating toolkit for anyone running Codex at scale right now — closer to a status page than a changelog, and more reliable than guessing from inside the product.

What to do about it

• If Codex is wired into an automated workflow (CI, a scheduled n8n pipeline, anything unattended), add retry/backoff logic now instead of assuming a fixed weekly cap through November 1.
• Check a reset tracker or Sottiaux's X account before kicking off a large batch job; resets have clustered in the morning Pacific time.
• Don't renew or upgrade a Codex plan on the strength of this month's resets alone. The pledge ends November 1, and nothing in it promises the cadence continues.
• If Claude Code's limits fit your workload better right now, say so and switch for that workload. Teams in the Stacked community have been comparing notes on exactly this tradeoff.

The short version

Twenty-one days into a 28-day pledge, OpenAI's Codex team has shipped 11 usage resets against 10 real improvements, and its own users are voting resets the clear winner — 80-90% approval against single digits for some features. The pledge ends November 1. What ships after that is the actual signal.