Nought for Three
Both models I called for November have now shipped, and I got neither date. I said GPT-6.2 Astra and Sol would arrive before 13 November, and OpenAI released them on the 17th, four days past the deadline. I said the next Fable would land in the last week of the month, and Anthropic put out Fable 5.2 on the 9th, a fortnight before my window opened. When the OpenAI call ran out I made a second one, for the week of 23 November, and the 17th was six days too early for that.
For OpenAI I took a 41-day median from ArtificialWatch’s release log and counted forward from GPT-6.1 Sol, which gave 9 November. I should have looked at what the median was made of. The log had three OpenAI models in it, so two gaps, 56 days and 26, and 41 is simply halfway between them. Worse, I counted from a Sol while waiting for an Astra. In the same post I’d told myself that Anthropic’s top tier runs slower than its cheaper ones, and then didn’t apply that to OpenAI, where two Sols shipped a week apart in September. There was no Astra-to-Astra gap to count at all until this week, and the first one, GPT-6 Astra to GPT-6.2 Astra, turned out to be 75 days.
The Anthropic call had the right tier and even less evidence. The Fable tier had exactly one gap on record and I treated it as a schedule. I made it 84 days by counting from Fable 5’s announcement on 9 June, but that release was suspended and redeployed on 1 July, and ArtificialWatch dates the model from the redeployment. Counted that way the gap is 62 days, which would have pointed at 2 November, a week from the real date. The gap that followed was 69.
Both releases landed just ahead of Anthropic’s listing. Its shares are due to start trading tomorrow, and Fable 5.2 arrived ten days before that, GPT-6.2 two. Last week I wrote that I’d want my flagship in the news that week if I were Sam Altman, and then guessed OpenAI would hold off until the IPO headlines had passed, so I had the listing in view and still picked the wrong side of it.
Each lab now has one flagship gap on record this autumn, 69 days for Fable and 75 for Astra. Add them on and both next flagships fall in the second half of January, which is the same sum I did for Fable on the 6th, with one gap and no reason to trust it, so I’m not calling January.
Sources:
-
Introducing GPT-6.2 Astra and Sol — OpenAI
-
Claude Fable 5.2 and Claude Mythos 5.2 — Anthropic
-
Anthropic shipped 6 of the 14 frontier models since July — ArtificialWatch
-
OpenAI research and release index — OpenAI
-
Model announcements — Anthropic
-
Claude Fable 5.1, Anthropic’s new frontier model is now available on AWS — AWS
-
OpenAI undercuts Claude by 95% with GPT-6.2 Astra in Anthropic’s IPO week — Dealroom
Filed under AI & machine learning
This post is timestamped using Blockchain technology. Verify