LatentMathBench: Investigating Latent Reasoning in Astra

After hearing rumors that GPT-6 Astra uses some form of recurrent depth, I decided to make a benchmark that specifically forces LLMs to do long chains of reasoning in latent space, without using chain-of-thought. The results for Astra are crazy: it can do 34 consecutive simple math operations (add, subtract, multiply and divide, all with small numbers) without using chain-of-thought! For comparison, Sol can only do 8!

I tested 24 other models (also from Anthropic) and none got anywhere close to Astra's result. What's even stranger is that there has been barely any progress on this benchmark (the second best result was Claude Opus 4.6, with 12 steps), until the release of Astra.

The article has the full results, along with some speculation about what might be happening inside Astra that could explain this result.