OpenAI's 28 days of Codex improvements: so far the race is over speed and allowance, not new models

Why it matters

【Key points】

Tibo (@thsottiaux, bio "Codex & ChatGPT @OpenAI") pledged (Oct 5, Beijing time) that for 28 days OpenAI would "either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset." Earlier he limited work to "simplifications, more efficiency for more usage, groundbreaking features or new models." Days 1–3 brought ~50% faster default speed, auto-review off the usage meter and a "banked reset" for paid accounts. GPT-6 in ChatGPT and the Decisions API we covered separately. Editorial synthesis, mainly from his X posts.

【Verifiable】

Pledge: https://x.com/thsottiaux/status/2106845241357824205 (~7.84M views, 26K likes shown on X at writing). Day 1: GPT-6 Astra and GPT-6.1 Sol "~50% faster" on subscriptions, incl. partners using Sign in With ChatGPT (OpenCode, Pi, Amp, Devin): https://x.com/thsottiaux/status/2107158998495748264 . Day 2: auto-review "does not use usage" (2–10% of plan, he says); simplified API; meeting notes; Decisions API: https://x.com/thsottiaux/status/2107575657014468879 . Day 3: GPT-6 in Chat; "40M active users across Codex and ChatGPT Work"; banked reset: https://x.com/thsottiaux/status/2107913674593644711 . The 40M is from a personal post, with no time window; not a company filing. "Reset" is undefined; users read it as a usage reset.

【Our read】

1. So far Codex users got speed and allowance, not a new coding model; with reset as fallback, usage itself is run as a product.

2. Via Sign in With ChatGPT the speedup reaches OpenCode, Devin and others: the subscription is becoming a channel for coding tools.

3. "Stalling for a new model" is Chinese X speculation, contrasting Anthropic's Opus/Sonnet/Haiku 5.5 run, not fact; new models are on Tibo's list.

4. Test: over the remaining 25 days, tally new-model or breakthrough days vs speed, allowance and reset days. If the latter dominate, the fight is usage economics; a new coding model would undercut the stalling thesis, not the channel one.