GPT-6 Astra can do a lot of multi-hop reasoning without chain of thought

See this post on LessWrong here.