OpenAI and Anthropic have begun limited gray testing of GPT-6 Sol and Opus 5.2, days after CEOs Sam Altman and Dario Amodei called for slower frontier-model development over safety concerns. Neither model has been officially announced, but selected users have reported access through random routing. GPT-6 Sol is described as a lower-cost model retaining GPT-6 Astra's performance, while Opus 5.2 is being tested inside Claude Code with reported gains in frontend work, SVG animation and complex 3D modeling. The simultaneous testing has fueled speculation that the companies could unveil broader model lineups this week. Analysts argue the slowdown rhetoric may also reflect a shift from costly training toward revenue-generating inference ahead of potential public-market events. Meritz Securities estimates OpenAI used more than 100,000 Nvidia Blackwell-class GPUs to pretrain GPT-6 Astra, creating an estimated $1.21 billion opportunity cost over 10 weeks, or about $6.3 billion annualized. OpenAI has separately announced the retirement of GPT-5.5 on October 14.