zenith looks like openai’s flagship gpt-5 model, fast, strong reasoning, and hit a perfect simplebench score, clearly a high-performance moE model. summit is probably the same model but set up for deeper reasoning, slower but smarter, consistent with openai’s “dial-a-depth” rumours. lobster feels like a smaller, coding-specialised “mini” gpt-5 variant, great for games and coding tasks. all three share the gpt-4o tokenizer, strongly hinting they’re official openai models, not open-source variants.