Today is 71 days since GPT-6 Astra went out on 3 September, and OpenAI still hasn’t shipped another flagship. My own call was GPT-6.2 Astra and Sol before 13 November, so this is the day it runs out. I got there from OpenAI’s 41-day median gap between frontier models, which is the same mistake I caught myself making with Anthropic in that post: the median comes from a burst across cheaper tiers, and the top tier runs slower. The nearest thing to a launch this week was a Codex lead musing about “a banked reset” on Thursday, which OrcaRouter noticed was quote-posted as a GPT-6.2 notice within four minutes. It was a quota coupon.

Below the flagship, OpenAI hasn’t been idle. GPT-6 Sol and Luna landed on 22 September, GPT-6.1 Terra on 26 October at a tenth of Astra’s price per token, and GPT-6 Cyber on 5 November, so anyone paying for ChatGPT or the API has had something new every few weeks. daily.dev quoted a tracker saying a miss last week would make this OpenAI’s longest gap between frontier models since GPT-5 and GPT-5.1, which says more about how quick the cadence had become than about any stall. The impatience lives with the people who watch Codex analytics for stray model labels. I’m one of them, and I’ve refreshed OpenAI’s changelog more this week than I’ve ever refreshed a parcel tracker.

Investors are a different matter. OpenAI raised $11bn at a $1.25 trillion valuation last week, short of the $20bn it first tried to raise, and Newcomer’s Tom Dotan wrote that OpenAI is “thirsting for a comeback story”. The same TechCrunch piece has Anthropic pricing its shares at $70 to $78, with trading reported for around 19 November and Fable 5.2 already out. I’d want my flagship in the news that week if I were Sam Altman. Still, the people who put money in at $1.25 trillion are betting on years, not on a launch date. A model that arrives a week after Anthropic’s listing barely dents that bet. A bad one would.

September already showed what a bad one looks like. The follow-up to the shipped model, GPT-6.1 Astra, was scrapped before release because, as Saachi Jain told WIRED, “it didn’t quite meet the bar in terms of staying within scope and authorization”, or in how it reported its own work back to the user. I wrote last week that built isn’t the same as cleared. An Astra pushed out to beat a rival’s listing, which then misbehaved in public, would cost OpenAI the one argument it has made all autumn, that it holds models back when they aren’t ready.

Prediction-market traders, as Mantis reports them, expect the next Astra in late November and treat a release by the end of the month as all but certain. My new guess is the week of 23 November, deliberately just after Anthropic’s first days of trading, so OpenAI avoids landing under the IPO headlines and gets the following week to itself. The real limit is later, around mid-December. By then Anthropic will have its first weeks as a listed company behind it, every number it reports will be set beside OpenAI’s, and a gap of more than three months at the top tier will need an explanation OpenAI hasn’t given. I’ll hold this guess more loosely than the last one.

Sources: