Monday was OpenAI's big DevDay thing, and by now you've probably read six different recaps of it already, so I'm not going to sit here and list every announcement like a press release. Custom GPTs, the Assistants API, the store they're promising for later, whatever. Half of it feels like it exists so there's something new to demo on stage. What actually made me sit up was two boring lines in the pricing table: GPT-4 Turbo now handles 128,000 tokens of context, and the input price dropped to a cent per thousand tokens. That's roughly a third of what old GPT-4 cost.
A cent a thousand doesn't sound like much until you realize what that context window means in practice. 128k tokens is something like 300 pages of text. So naturally the first thing I did Tuesday night, instead of anything productive, was point the API at my own archive and see how much of this blog I could cram in at once.
I didn't do the whole twelve years, that would've blown past the limit and my patience both. I pulled everything from January 2022 through last month, exported the posts to plain text, and ran a rough token count with tiktoken. Came out to just under 96,000 tokens for about 340 posts. Fed the whole thing in as one blob with a prompt asking it to summarize recurring themes and call out anything I'd contradicted myself on. Total input cost for that one call: about ninety-six cents. Under six months ago the same call on GPT-4's 8k model would've been impossible without chunking it into a dozen separate requests and stitching the summaries back together, and it still would've cost more.
The summary itself was fine, honestly kind of a letdown after the buildup. It correctly noticed I complain about build tooling more than any reasonable person should, and it flagged that I said in early 2022 I was "done with React for personal projects" and then wrote about a React side project four months later. Fair catch. But nothing revelatory came out of it, no hidden insight about my own writing I didn't already know. Which I think is the honest version of what a 128k context window gets you right now — not magic, just less duct tape. You stop babysitting chunking logic and start just pasting things in.
The other thing from this week I want to flag, mostly as a complaint: xAI's Grok. It went out to a small slice of X Premium+ subscribers a few days before DevDay, and everything I've read about it leans hard on the "it has a rebellious streak" pitch, like sass is a feature. I get that it's early and invite-only and I haven't touched it myself since I'm not paying sixteen bucks a month for the privilege of maybe getting sarcasm from a chatbot. But branding a model's personality around being edgy feels like the kind of thing that sounds fun in a pitch meeting and gets walked back within two product cycles once people actually use it for something that matters. We'll see. I'm not betting against it existing in some form a year from now, I'm just skeptical the "rebellious AI" angle survives contact with actual users asking it to help debug a cron job.
Back to the context window thing for a second, because it's the part I actually think matters longer term for anyone building on this stuff. The chunk-and-summarize dance has been the default workaround for anything document-shaped since these APIs launched — split your text, run it through separately, merge the results, hope nothing important got sliced across a boundary. 128k tokens doesn't kill that pattern for genuinely large corpora, but it moves the threshold for "genuinely large" a long way out. My whole blog archive minus a couple years now fits in one shot. A decent chunk of a codebase fits in one shot. A full legal contract fits in one shot. That's a different kind of tool than what existed in October.
I still owe a longer post on the Assistants API once I've actually built something with it instead of poking at it for twenty minutes. For now I've got a summary of my own writing habits that cost less than a Tuesday-night coffee to produce, and honestly that mundane fact is the more interesting story than anything that happened on stage Monday.