EpochAI also released some estimates for GPT-6 compute and params

I think total training compute estimate is fine, but 5.35x more pre-training compute than RL compute seems silly

the median for the pretraining duration is also very high with a 90% interval of 42–135 days

the most obvious flaw is that for GPT-5.5 (released on April 23) we know that it was trained on Stargate Abilene, and other models like GPT-5.3-Codex, GPT-5.4, and the 5.6 family were probably trained there too, which means that the pre-training of Astra couldn't have used all 100k GPUs on average in all scenarios they model

OR more simply, pre-training just happened later than they modeled after GPT-5.5 release (I put it at early May - late July)

for the model size estimate they also don't really consider trends in sparsity or looping

lets not talk about the param estimates. they are obviously way too high:
- "Across the named scenarios: active 0.1T-4.8T; total 1T-143T (corners: p5 FLOP, 400T, x10 -> p95 FLOP, 50T, x30). Central case (p50 FLOP, 100T tokens, sparsity 20): active 1.2T, total 25T."