@catboy1006: In a single week, the three major labs turned "launches" into three kinds of performance art: OpenAI skipped 6.1, and GPT-6.2 Astra quietly began a gray rollout into Codex — 19.8% of traffic in billing has already switched to the new base model, with no change to the frontend name. Apparently 6.1 failed safety testing and went back to the drawing board for a month before skipping straight ahead; Codex's underlying…

In the span of a week, three major labs turned “releases” into three different forms of performance art:

OpenAI skipped 6.1 — GPT-6.2 Astra has quietly gone into staged rollout on Codex. According to billing data, 19.8% of traffic has already been switched to the new backbone, and the frontend name hasn’t even changed. Apparently 6.1 failed safety testing, was sent back for a month, and got fast-tracked straight to a skip-level release. The Codex underpinnings were rewritten from the ground up in 28 days, enabling it to spawn its own subprocesses and run continuously for hours.

Anthropic’s move is even more interesting: Claude Mythos 5.5 is sitting right there in the website’s source code. The same brain is split behind two gates — the unguarded Mythos is only offered to top cybersecurity and research institutions, while Fable 5.5 with a 1 million token context has already surfaced on Azure. Leaks claim Mythos broke into the Linux kernel and obtained root access in 53 minutes.

Google, meanwhile, has locked away the leaderboard: Gemini 4 Carbon topped the chart at 59 points, but it’s only available to vetted cybersecurity teams.

When “can you actually get access to it” is scarcer than “how high does it score,” looking at rankings matters less than looking at access policies.

#AI #OpenAI #Claude #Gemini #LLM