The new AI model uses diffusion reasoning to generate 1,000 tokens per second; it runs about 5x faster than Haiku, speed limits are ...