The Best Deal in AI? GPT-6.1 Sol Is 5× Cheaper Than Astra

We are launching GPT 6.1 Soul. It gives you very near Astra level intelligence at a fifth of the price. >> On the same real jobs with the same right answers, the flagship GPT 6 Astra cost $2.69. GPT 6.1 Soul, $0.51. So, should you switch? Three things first. A rival at the exact same price finished the same work six

times faster. Telling Soul to think harder made it worse at coding. And every OpenAI model we tried wrote the same catchphrase. GPT 6.1 Soul launched this morning at OpenAI's Dev Day, [music] 1 week after GPT 6 Soul, and that one had a rough week. On Hacker News, one developer said it was worse than the model before it,

far worse. So, [music] where does 6.1 fit? OpenAI now sells three sizes, price per million tokens, which is roughly 750,000 words. Astra, the flagship, $10 to read a million, 50 to write them. Luna, the little one, 10 cents and 50 cents. And Soul in the middle, $2.10. Here's the first surprise. The price of

6.1 Soul didn't change. It's exactly last week's price. What changed is re-reading. When a coding agent works on your project, it sends the same big pile of context, your files and its instructions, [music] again on every single step. OpenAI keeps a copy of that pile, a cache, and charges less to read it again.

That re-reading price just halved, >> [music] >> from 20 cents per million tokens to 10. And it adds up. In our own runs in Codex, about nine of every 10 tokens Soul read were re-reads. One Hacker News regular called the cheaper cache the actual big announcement. Where can you use it? In ChatGPT work

and Codex on paid plans, [music] and in the API. Not in regular ChatGPT chat, yet. And update Codex first. The version on our Mac flat-out refused the new model. >> [music] >> The update that came out this morning runs it. There's a faster version, too. Ultrafast is out today for Astra. >> [music]