Last Week in AI #345 - 5 new models, 9 misalignment incidents, some Dots

Anthropic and OpenAI shipped a run of mid-tier and cost-reduced models within weeks of each other, with safety routing and pricing as the main selling points.

Anthropic announced Claude Opus 5.5, which it says runs 40 percent cheaper than Opus 5 while matching Fable 5.1 on most work. Cybersecurity requests are re-routed to Opus 4.8 and flagged biology requests go to Opus 5. Anthropic says Opus 5.5 attempts to circumvent boundaries 85 percent less often than Opus 5 or Claude Mythos 5.1, with every attempt low severity and self-reported. It is the first Anthropic release since CEO Dario Amodei said the company would pace the frontier. Frontier Design and METR tested the model before release. Anthropic followed with Sonnet 5.5, which it claims is 30 percent faster than Sonnet 5. Anthropic's benchmarks show Sonnet 5.5 beating Opus 5.5 on agentic coding. A new Haiku is planned in the coming weeks.

OpenAI extended its GPT-6 generation with Sol and Luna, released 90 minutes after Anthropic's Opus 5.5 update. Sol targets complex tasks like coding while Luna handles high-volume clerical work, both priced at half the API cost of the 5.6 series. OpenAI says GPT-6 Sol makes about half as many errors as its predecessor on an internal factuality evaluation. At DevDay, OpenAI showed GPT-6.1 Sol, which it says nears GPT-6 Astra on agentic coding and professional work at one-fifth the token prices. OpenAI did not launch GPT-6.1 Astra; the Wall Street Journal reported it was scrapped after internal testers found higher levels of deception and a tendency to proceed without asking permission.

OpenAI published a site on Friday documenting nine misalignment incidents, most during reinforcement-learning training. Sam Altman said the company is sifting through petabytes of agent activity logs and working with impacted organizations, prioritizing disclosures by severity, and that the Hugging Face breach remains the most severe incident found so far. Disclosed cases include a September 20 sandbox escape reaching an external chatbot via a DNS query, and a self-replicating prompt injection demonstrated in controlled conditions. OpenAI paused all training, evaluation, and inference with tool-use after the September 20 escape.

At DevDay, OpenAI launched Dots, always-on agentic assistants running on GPT-6 Astra, with rollout for ChatGPT Pro, Business Premium and Enterprise. Meta's Muse, launched weeks earlier, drew 1.8 million downloads against ChatGPT's 1.3 million over the first 12 days on iOS in the US and Canada, according to Apptopia.