OpenAI dropped a new model lineup yesterday, GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, straight into the API, no ChatGPT rollout drama, no waitlist. I almost missed it because it happened the same week everyone's been glued to tariff news, but a friend on a Discord I lurk in pinged me at like 11pm going "did you see the context window on this thing" and that got my attention fast enough that I stayed up messing with it instead of sleeping.
The headline spec is a million-token context window. GPT-4o topped out around 128k, which already felt generous until you actually tried to stuff a whole codebase into it and watched it start forgetting the function you defined three files ago. A million tokens is a different category of thing. That's something like 700,000 words if you're doing rough English-text math, which is more than the entire Lord of the Rings trilogy with room to spare. I don't have a project that needs that much space, honestly, but I wanted to see what breaks first: my wallet or the model's attention.
So I did the dumb obvious test. I took the full markdown export of this blog (every post since November 2011, which is honestly a humbling amount of text to look at all at once) and dumped the whole thing into gpt-4.1-nano with a prompt asking it to find posts where I contradicted myself about something. Not a serious research project, just curiosity. It found three, including one where I apparently declared RSS dead in 2013 and then wrote a whole nostalgic post about switching feed readers in 2019. Fair catch, honestly. I'd forgotten I did that.
The pricing on nano is the part that actually matters for someone like me who runs personal scripts and not a startup. It's cheap enough that running my entire blog archive through it cost less than a coffee, and that's with a model that's apparently scoring better on coding benchmarks than GPT-4o did a year ago. The full GPT-4.1 model is pricier, somewhere around two dollars per million input tokens, but even that's not scary money unless you're doing this at real scale. Mini sits in between. The naming is doing the same thing OpenAI naming always does, where you need a spreadsheet to remember which one is smart-and-slow versus fast-and-cheap, and I say that as someone who has now made that spreadsheet twice for two different model generations. I wish they'd just pick adjectives instead of decimal points.
One thing buried in the announcement that I think is more interesting than the headline number: they're deprecating GPT-4.5 preview, the giant expensive one they shipped back in February, and pulling it from the API in July. That model barely got a chance to exist. It came out, everyone said "huh, expensive, not obviously better," and now it's basically already being retired in favor of 4.1. I don't know what that says about how fast this stuff is moving except that it's fast enough that a model can go from launch to legacy status in under five months, which is not a pace any other kind of software ships at. My web host still runs a control panel that looks like it's from 2009.
If you're building anything that chews through long documents, legal contracts, log files, an entire git repo for code review, this is worth an afternoon of poking at, especially on the nano tier where mistakes cost basically nothing. If you're just chatting with a model day to day, you probably won't notice much difference from 4o, and it's not even in the regular ChatGPT app yet as far as I can tell, just the API and the developer playground.
I keep telling myself I'm going to build something real with one of these long-context models instead of just running party tricks against my own archive, and I keep not doing it because the party trick is more fun than the actual planning work. Maybe next week I point it at my email instead and regret it immediately. I'll report back if the model finds anything more embarrassing than the RSS thing.