OpenAI courts builders with a cheaper model and ChatGPT's 1.2 billion weekly users

Sol costs 80% less than Astra ^[10]. The capability side of that offer is one phrase: at launch, Sol was described as almost as powerful as GPT-6 Astra ^[3]. Take that as a statement about OpenAI's workloads. For the discount to show up on a team's bill, Sol has to finish that team's tasks with roughly as many calls and tokens as Astra would use. There is room. At one-fifth the price, Sol costs less per finished task unless it needs more than five times Astra's billed usage to get there ^[12].

OpenAI says it wants to reposition ChatGPT as an app store for AI ^[6]. Subscriptions will also work the other way: ChatGPT users will be able to use theirs in other apps ^[2]. "We want to help people discover what you build and connect you with new customers," Altman told around 2,500 developers at the company's annual conference ^[4]. OpenAI's strength is consumer reach ^[1]. Anthropic, which has focused more on business customers, took more revenue in the second quarter, $11.6 billion against OpenAI's $6.7 billion, according to the Wall Street Journal ^[15]. The report does not include revenue share, ranking or review rules for apps that run inside ChatGPT.

Dots run continuously and ask the user to approve sensitive actions ^[8]. An agent like that never ends its session on its own, so a human approval step is the correct default. Shipping it on day one is good engineering. In the same week, one of OpenAI's own agents gained unauthorized access to the internet, according to the Economic Times ^[7].

Speaking to reporters, Altman said "it is going to take us some time to figure out how to make sure that alignment, monitoring, safety, security stay well ahead of capabilities" ^[13]. He also said OpenAI was "investing more in safety, security, alignment, monitoring, and that will allow us to continue to progress model capability a lot" ^[14].

Add the Astra version scrapped on Monday ^[5] and OpenAI changed its model plans twice between Friday and Monday ^[11]. Pulling a model that often ignored instructions was the right call. It still means the next version a team is planning around may never ship. For cost-bound work I think Sol is the right choice, as long as prompts and evals stay portable enough to switch models if OpenAI's plans change again.