GPT-6 Astra's control over its reasoning traces makes chain-of-thought monitoring harder
GPT-6 Astra scored 60.9% on a reasoning-length control test versus 16.1% for GPT-5.6 Sol, suggesting monitors that read chain of thought for cheating are losing their edge.
I think it’s quite funny that models being bad at controlling their own written reasoning might’ve actually been protecting us. For context; I found out that one of the ways labs monitor AI is by having another model read its chain of thought for signs of cheating or suspicious behaviour. And back in March, OpenAI found that 13 reasoning models were surprisingly bad at controlling what appeared in those traces. Success rates ranged from just 0.1% to 15.4%. So OpenAI considered that a good news. But that weakness isn’t something labs can count on forever. Because now GPT6 ASTRA scored 60.9% on a comparable length reasoning test, against GPT5.6 Sol’s 16.1%. And when explicitly told to conceal underperformance, Astra also became harder for certain monitors to catch.  This also goes beyond OpenAI. Anthropic found that Claude and DeepSeek R1 often failed to mention information that influenced their answers. Which also gives us yet another reason we can’t assume a model’s written reasoning tells us everything. The FMF has also warned that latent reasoning could move more computation into internal states that chain of thought monitors can’t read. And I think this is a problem. We’re basically getting better at building models that reason, but way behind in figuring out how much we can trust the reasoning they show us. Are we cooked?