I think we can see indirect evidence of whatever changes OpenAI made to GPT 6’s architecture in its success in handwriting. GPT-6 Astra is the first reasoning model where transcription accuracy for handwriting actually improves with more thinking/effort. This might sound counterintuitive, but since O1 it’s been the case across Gemini, Claude, and Grok. My intuition is that these models could not improve visual reasoning through effort as it often made them second guess their initial judgements (we saw this with the clock tests too). So the best results on handwriting always came from minimal thinking. Looks like Astra is indeed doing something new inside because that trend has now reversed which ai think is a significant signal far more important than the idiosyncratic nature of historical handwriting might suggest. Deciphering historical handwriting is a visual task to about 95% accuracy and then above that, it increasingly becomes a reasoning/logic task: using a combination of document context, outside knowledge, and deductive logic to get to 100% (which is unobtainable even for humans). It requires symbolic reasoning and a world model (ie the past is composed of different elements than the present yet can coexist with the present) to get there. We saw glimmers of this last year with Gemini 3, but this is a real signal that the models have started to crack symbolic reasoning. On high, it scores 2% modified word error rate (WER) and 4% character error rate (CER) — where lower is better. On low it scores 2.41% modified WER and 4.74% CER. Interesting times!