
some thoughts on claude mythos 1) claude mythos is expected to be a model tier above opus, both more expensive and with greater capabilities 2) the model is supposedly going to be rolled out to a small number of early customers with a focus on cybersecurity 3) the model is going to be high risk or critical risk equivalent for cybersecurity 4) i believe this makes it the first model to really begin to show how labs will behave around models with actually dangerous capabilities 5) in addition, mythos is likely one of what will be a number of much larger models being trained and released over the next year 6) i think there has been some trepidation to train larger models at least since the commercial failure of gpt-4.5 7) however, there is 8x as much compute available in the world today as at the start of 2024 so there is much more capacity 8) and scaling is not dead and larger models are probably going to be more sample efficient in rl so i think we should expect more reasoning gains 9) also though, once these capabilities exist, they tend to be very distillable, so we should also expect very capable small models to follow (like o3 to o3-mini) 10) i think it is going to be a wild 2026; so much more compute will come online fortune.com/2026/03/26/ant…


















