Some extended thoughts on GPT-5 1) GPT-5 is a good model. It feels like it provides better search and performance than o3 did before it. 2) It's disappointing to people because it is an incremental improvement, which does not open up fundamentally new use cases. 3) The really interesting story around GPT-5 seems to be more about competition with Anthropic. 4) Anthropic has increased its revenue by 4-5x over the last 6 months. OpenAI has increased it's revenue by 2x. 5) A lot of Anthropic's revenue growth is due to API revenue, which is a much larger percentage of Anthropic's revenue (60%) than OpenAI's revenue (25%). 6) About 50% of Anthropic's API revenue comes from its Cursor and Github Copilot partners. Anthropic probably collects something like $800m in revenue from Cursor. 7) GPT-5 seems to be in part about challenging Anthropic's dominance in coding agents. GPT-5 is now the default in Cursor. 8) GPT-5 finally matches Claude 4.1 Opus' performance on SWE Bench Verified, which isn't a perfect measure but which seems to be a good proxy for performance. 9) The cursor partnership has the opportunity to steer a lot of revenue away from Anthropic, while helping OpenAI to cement its share of the consumer mind for coding applications. 10) I wouldn't read too much into it, but it could slow Anthropic's revenue growth and make it marginally harder for them to raise (although, probably a weak effect). 11) I have other thoughts around GPT-5 from a user interaction / launch perspective. 12) I think they botched the launch; no one wants to watch live streams, the benchmarks are not intelligible anymore, and there was nothing viral to interact with. 13) Cool model interactions need to be about new modalities or need to be very agentic, which requires a lot of scaffolding. World historic stuff is good too - our models solved a millennium prize. 14) Labs tend to solve UI/UX in one place and then just end up with problems in another; we simplify the model complexity (somewhat) but now have to pick personalities. 15) But, this is okay, and just part of the grand adventure where we work toward truly tailored interactions with our LLM assistants.