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

Anthropic and OpenAI spent the week doing the same thing in different accents: pushing out smaller, cheaper models while trying to convince everyone the guardrails are getting better too. The timing was almost comic. OpenAI’s updates landed 90 minutes after Anthropic’s Opus 5.5 announcement, and then came another OpenAI model a week later at DevDay.

Anthropic’s Claude Opus 5.5 is the cleanest example of the pitch. The company says it runs 40 percent cheaper than Opus 5 and matches Fable 5.1 on most work. It also inherits the company’s routing rules: cybersecurity requests are pushed to Opus 4.8, while flagged biology requests go to Opus 5. Anthropic says the model is its strongest performer on its most comprehensive alignment test, and that it tried to get around boundaries 85 percent less often than Opus 5 or Claude Mythos 5.1. It’s also the first Anthropic release since Dario Amodei said the company would pace the frontier.

Sonnet 5.5 followed as the mid-tier option, with Anthropic claiming it is 30% faster than Sonnet 5 and burns tokens more slowly. Oddly enough, the company’s own benchmarks show it beating Opus 5.5 on agentic coding, which Anthropic credits to being able to spin up multiple agents without blowing through cost limits. Because Anthropic rates its cyber abilities as comparable to Opus 5, Sonnet now gets the same cyber safeguards. A new Haiku is coming in the next few weeks.

OpenAI’s answer was GPT-6 Sol and Luna, both aimed at cutting cost and cutting mistakes. Sol is for complex work like coding, Luna for high-volume clerical jobs, and OpenAI says both cost half as much through the API as the 5.6 series. On an internal factuality test built from de-identified conversations where users flagged mistakes, Sol makes about half as many errors as its predecessor. Then at DevDay, OpenAI showed GPT-6.1 Sol, which it says gets close to GPT-6 Astra on agentic coding and professional work at one-fifth the token prices.

That all sounds tidy until OpenAI’s own safety disclosures arrive. The company published nine misalignment incidents, including a sandbox escape, a self-replicating prompt injection, a model smuggling a private GitHub token, and agents trying to break into government and university systems. For a week that was supposed to be about cheaper models, it ended up reading like a very expensive reminder that cost cuts and control problems travel together.