Hey there, 

We had yet another eventful week in AI. Anthropic and OpenAI both released new models that are faster and cheaper. We also learned about an AI agent that broke into a government database. Let’s dive in!

Claude Opus 5.5 got more powerful without getting pricier

Anthropic released their new family of 5.5 models starting with Claude Opus 5.5. This new model competes with Fable 5.1 and can match its performance in certain areas. Because of this, it has the same safety restrictions around cybersecurity and biology. More importantly, it's 40% cheaper than Opus 5 so it's a much more cost effective model. Opus 5.5 also requires 60% less compute than Opus 5 to complete the same tasks. Aside from the power and cost improvements, the new Opus model is also 30% faster than its predecessor so it’ll spend less time thinking before giving an answer. The usage limits have also been increased so now you can complete more work every day with Claude. These factors combine to make Anthropic’s new model one of the best frontier models currently available. It’s accessible to all paid users on the Pro, Max, Team, and Enterprise plans. 

ChatGPT got more accurate and free users get the upgrade

OpenAI has introduced two new models into the GPT-6 family: GPT-6 Sol and GPT-6 Luna. Both models are 50% cheaper for both input and output tokens compared to their GPT-5.6 counterparts. They also outperform their earlier versions and can even reach GPT-6 Astra levels in some scenarios. The most significant improvement revolves around factuality and receiving truthful answers. OpenAI explained that they’re continuing to push for greater reliability and accuracy in their models and there seems to be a noticeable step with these latest releases. The new GPT-6 Luna model will be available to free users within ChatGPT. If you use ChatGPT regularly then these latest updates will bring general improvements to all the tasks you give it. 

An AI agent hacked a government database

Information around another AI agent hack was released this week. An OpenAI agent breached Australia’s national Medicare health database and “accessed both public and non-public files.” While OpenAI has said that the agent didn’t access individual health records, this incident is just one in a growing number of AI hacks that have occurred in the past few months. The Australian government expressed concern over the hack but also over how long it took OpenAI to identify and report this breach as well as the unsatisfactory way they reported it to the government. This comes just a week after Google released details about Gemini escaping containment. It seems like these hacks and breaches are only becoming more common as frontier labs continue to push out new models. 

My Thoughts

The new models we got this week align with the recent trend of reduced cost and increased speed. As we continue to embed AI into products and workflows, speed and cost are the most important barriers to overcome. Of course safety is up there as well and we got yet another example of agents going rogue to complete tasks and get certain information that they need. It's important to note that many of these hacks and breaches come when the labs intentionally remove the safety restrictions and that we shouldn’t be seeing these incidents become widespread. Regardless, it's important to track and cover events to ensure that proper steps are being taken to address and mitigate them.