[00:00] Last week, Sam Altman sat frustrated in his bunker as he watched Zuck unveil Muse, a free personal assistant-inspired spy that trains on everything you tell it in exchange for doing your shopping. Not to be outdone, Sam got on a stage of his own in San Francisco just a few days ago
[00:15] and announced that for $100 per month, they'll give you another cute and totally harmless personal assistant that can respond to all of those AI-generated emails in your inbox and rewrite your backend in whatever language DHH thinks is fast that day.
[00:28] In today's video, we'll break down everything else you missed at OpenAI Dev Day, including a model that's almost as smart as GPT-6 Astra, an API that looks suspiciously familiar, and a new subscription tier that quietly castrated the $200 per month plan.
[00:44] It is October 1st, 2026, and we're watching the Code Report. Back in January, a bored and recently retired Austrian developer named Peter Steinberger released QuadBot, a first-of-its-kind AI personal assistant for people who like to appear productive
[00:58] without the hassle of actually doing the work. It quickly became the fastest growing repo in GitHub history, which was enough to get him hired by OpenAI to continue to work on the idea. Then, just last week, all his hard work finally paid off
[01:10] when Zuck took the concept and stuffed it into Muse. And just a few days ago, his own employer did the same thing, but now with a $100 per month price tag that they're calling DOTS, with a singular dot being a personal agent powered by GPT Astra So it basically just a bot with the B flipped around You just give it a name then its cute swishy little avatar will get to work ruining the internet as we know it In their keynote demo Sam told his dot Alfred to remove an old inventory
[01:36] API before it got shut down. So Alfred then followed modern software best practices of tracing dependencies, updating integrations, running tests, and expanding the repo with three PRs from the CEO that some poor dev will feel obligated to review. Because they know it'll
[01:52] be the first time human eyes have ever looked at that code. Then, Holly Lee came out to prove that DOTS work outside of a pre-recorded video, and instead just proved that live demos are still hard even in the age of flop. But my favorite part of DOTS is when the internet realized that DOTS.com
[02:07] is owned by some woman's clothing brand that died in 2014, and DOT.com is owned by XAI, and now redirects to Glossbox's landing page. It's the best domain break I've seen since whoever got
[02:19] their hands on Next.js.dev and Remix.dev a few years back. But if that sounds like your slot can of choice, you'll need a pro subscription, which OpenAI revealed some new pricing for. The new Pro 500 plan costs $500 per month and gets you 25 times the usual usage of ChatGPT+,
[02:36] along with access to UltraFast, which is a new speed tier that runs Astra at 300 tokens per second. In the API, UltraFast costs $60 per million tokens in and 300 out. And when Sam
[02:48] said those words out loud the audience of grown men moaned You know what It worth it Which apparently isn the first time he caused that reaction Meanwhile the plan went back on sale but had its usage dropped in half Luckily if you broke they announced GPT Sol
[03:03] which they're describing as near-astrointelligence for a fifth of the price. You get $2 per million tokens in and $10 out, which is the exact same price of GPT-6 Sol, which launched a week ago. Then, Sam announced the new Decisions API, where the idea is that instead of asking a big model to write a paragraph and then parsing the JSON out of it with a retry loop, you hand their tiny Luda model a fixed list of answers, and it picks one in a fraction of a second.
[03:29] For example, imagine you're building a dating app for gay doggies. As a way to fund the venture, every message needs to be categorized by how useful it would be for blackmail. Historically, you'd send each message to a full language model, wait a few seconds while it pretends to be ethical, and then hope the JSON comes back in the right shape.
[03:44] With a decision model, you give it a finite set of answers up front, it returns one of them in a few hundred milliseconds, and since it physically can't answer anything outside your list, there's nothing to parse or retry. And if all this sounds familiar, it should, because just two weeks ago, we released a video about a company called Typesafe AI, founded by an OpenAI researcher who just released Jev that does the exact same thing.
[04:06] but they didn't stop there. Later in the event, they announced a Google Docs clone called Pages, a Notion clone called Space and a Slackbot that works a lot like QuadTag showing what it looks like when you have infinite tokens But the one thing in the keynote that could actually make you money is sign in with ChatGPT which lets your users log into your app with their ChatGPT account
[04:26] and burn their own tokens instead of yours. It launched with 16 partners, including Devin, Notion, Vercel, and OpenClaw, and comes with an implicit understanding that if you get too popular, they'll just clone your app too. Which is why if you need to build a decision model into your own app,
[04:40] you might want to own the weights yourself. And you can do that with Fastena Labs, the sponsor of today's video. Jeb has been going viral for getting developers fast type decisions, but it's closed weight and only runs on the company's cloud.
[04:53] The Fastino's Gliner 2.5 to side is an open weight model that makes those same calls locally and up to eight times faster. So you can train your own Jeb-like models and own the weights. It comes from the Gliner family,
[05:05] which has been open source since 2024 with over 50 million downloads. I added their skill to open code in about five minutes, so my agent can route each developer request to a coding model or internal tool
[05:17] and pull out the repo, file path, and function names. Other developers have used it to run a browser agent 36 times cheaper than Jev, and Fastino just released a brand new model called Glide that beats Jev on accuracy across intent routing, fact checking,
[05:31] and hallucination checks by a pretty wide margin. You can host these models anywhere, but Fastino's inference API is the quickest way to replace your frontier model. Try it out for free with our code at the link below. This has been the Code Report. Thanks for watching, and I will see you in the next one.