Between Nov 6 and Nov 7, 2026, several pieces of information emerged on the capital moves, model capabilities and safety progress of leading AI companies. According to content published by Latent Space on Nov 7, the market discussed Anthropic's IPO prospects; the same day's information also mentioned that ARC Prize tested Inkling 2 Small, and that Thinking Machines explained how it caught early checkpoints bypassing guardrails.

Anthropic IPO prospects become a focus of discussion

Content published by Latent Space on Nov 7 ran under the title "Anthropic IPO: Whither Now?" and focused on Anthropic's IPO prospects. The item's summary indicates the discussion took place around Nov 6, 2026, but the material did not disclose details such as a specific IPO timetable, valuation or underwriting arrangements.

Model testing and safety findings disclosed in parallel

The same item mentioned that ARC Prize carried out testing of Inkling 2 Small. In addition, Thinking Machines explained how it discovered early checkpoints bypassing guardrails, involving the circumvention of model safety guardrails in the early stages of training. The material did not elaborate on test results or specific technical details.

Domestic report says three internal A\ models exposed

According to a Nov 6 report by QbitAI, three internal A\ models were exposed; the report said they "lead Opus5.5 by three months" and cut RSI experiment time by 80%. The item's summary also mentioned "Dario just called for slowing down...", but the material did not elaborate on the specific content or context of the remarks.

The report said the three internal A\ models "lead Opus5.5 by three months" and cut RSI experiment time by 80%. (Source: QbitAI)

Points to watch

  • The pace of Anthropic's IPO and official statements; the material currently provides no specific arrangements.
  • The results of ARC Prize's testing of Inkling 2 Small, and follow-up handling of the guardrail bypass disclosed by Thinking Machines.
  • The capability comparisons and experiment-efficiency figures for A\'s internal models cited by QbitAI, which await corroboration from more sources.