OpenAI's "supermassive black hole" AI model (4.1)

OpenAI introduces GPT 4.1. We also finally find out what Quasar is. Also, you might recall that OpenAI model, if the rumors are to be believed, will cost $20,000 per month and will resemble a PhD-level researcher. Apparently, that's the upcoming 03 and 04 mini, and it seems like the early previews are already being used to contribute to scientific discovery and coming up with novel ideas. And this is coming up very soon. Also, I feel like most people missed this. OpenAI did release their prompting guide. This is mainly for 4.1, although it can reply to more than just that model. But as you can see, some of these little tweaks in how we prompt models can have, for example, a 20% improvement in how it's able to do various tasks. But let's start at the beginning. So first and foremost, we have Quasar. So this was that stealth model that appeared in the LM System Arena. Sam Altman kind of hinted that this was one of the OpenAI models. And so a Quasar is a very energetic and distant active galactic nucleus powered by a supermassive black hole that emits exceptionally large amounts of energy across the electromagnetic spectrum. The OpenAI live stream where they announced GPT 4.1 was called, you know, developers and a little handshake emoji and supermassive black holes. I embarrassingly missed that reference. Okay, I get it now. So basically, it sounds like the Quasar model might have been a GPT 4.1 mini. I think they kind of let it slip in the live stream that that was the case. But it's alive now. And you're able to use it with the API. Or if you wanted to try it in a more kind of familiar format, you can use OpenAI Playground and test it out there. We'll do a more thorough testing. But here are the kind of like the big headlines about these series of models, this family of models, because they're releasing three different ones, the 4.1, 4.1 mini, and 4.1 nano. So for coding, 4.1 scores 54.6 on a SWE Bench Verified, improving 21.4% over GPT 4.0, make it the leading model for scoring. It's better than a GPT 4.5, but it's 54.6 on the SWE Bench Verified. SWE Bench is, of course, software engineering. And we're going to come back to that just a little bit because it's kind of interesting. It's important to understand what we need to be comparing this to. Do we compare this to Gemini 2.5 Pro? What is the best sort of apples to apples comparison? This model is supposed to be very good at coding. And that's showcased not just in the benchmarks, but also, you know, user testing in the LM arena. There's a quite a number of people that have came back and said that it's excellent for coding. There's also a long context, including in multimodal videos, basically, and instruction following. We'll come back to that in just a second. One thing they did mention during live stream is that developers, if you're using the APIs, you're able to get free OpenAI tokens if you're willing to share feedback and information with OpenAI. So you're able to share feedback with OpenAI and prompts, but also you're able to enable sharing prompts, completions, and traces with OpenAI. And this is where you get some extra tokens. Looks like you're eligible for free daily usage on traffic shared with OpenAI through April 30th. Get up to 1 million tokens per day for 4.5 preview, 4.1, 4.0, and 0.1, and up to 10 million on 4.1 mini, 4.1 nano, GPT-4.0 mini, and the rest of the mini models. So also you might have heard of Windsurf. So it's an IDE, an integrated development environment. So basically a tool that you use for coding, it looks like they're giving away free unlimited access to GPT-4.1. So as you can see here, so it's free from April 14th to April 21. This is not sponsored. I guess one of the founders actually came onto the OpenAI live stream and kind of announced this. So I actually plan to jump in and check it out. I've been using Cursor, but I haven't tried out Windsurf. So it looks like the GPT-4.1 is already integrated and ready to go. And you're also able to get free daily usage through OpenAI. Again, go to wherever your OpenAI API keys are generated. And in the bottom left, it'll have a little pop-up if you're interested in doing this. But let's kind of examine these three models that were introduced. You see them here in purple and the sort of the previous model, the 4.0, that's in green. So latency, so the slower it is, the more to the right on this graph it is. So 4.1 and 4.0 are roughly the same latency. So kind of like the same speed, 4.1 mini and 4.0 mini, again, roughly the same latency. And interestingly, the 4.1 nano, so this is the even smaller model than mini, is much faster, but the sort of the intelligence, which they're using the multilingual MMLU to kind of show the metric for that. So it kind of, yeah, it drops off, but there's probably some edge devices where this is just the type of thing you might need. But definitely a noticeable massive jump from 4.0 mini to 4.1 mini and a smaller but decent jump from 4.0 to 4.1. By the way, here's a great way to kind of wrap your mind around the difference between these models. So here we have the 4.1, the O3 mini and a GPT-4.0. So notice that 4.1, just like the 4.0, they are non-reasoning. They are the non-reasoning models. The O3 mini is a reasoning model. The O3 mini had a text input and output. Both the 4.0 and 4.1, they have a text and image input and text output, roughly the same speed across the board. But notice from the 4.0, you know, they sort of designate three sort of dots in terms of how smart it is. 4.1 and O3 mini are sort of four dots to signify that they're smarter than the 4.0. Basically what this whole thing is saying is they've sort of figured out how to make the non-reasoning models basically just as good as the reasoning models, right? Just as smart while still being as fast as like the mini version of the reasoning model and also able to do images as input. And as you can see here, you know, if we're looking at the 4.0 versus the 4.1, the pricing is reduced. So about a 20% reduction in the input and output. It's more expensive than the O3 mini. But as you can see here for coding, it beats out the other models, scoring 55 on the suite bench verified. You know, if we compare that to Gemini 2.5 Pro, which is currently kind of the reigning king, right? On the suite bench verified, it's getting 63.8. But keep in mind, that's the reasoning model. So it's not a perfect comparison. But I think the point is that these non-reasoning models are sort of catching up, getting a little bit closer to the top of the line reasoning models. On the 8-Ears Polyglot Benchmark, it does better than the 4.0. Not quite as good as the O1 high and the O3 mini high, but better than the GPT 4.5. The one thing they talked about quite a bit is the improved instruction following. GPT 4.1 follows instructions more reliably. Keep in mind, they posted that prompt guide that kind of goes along with this. They can show you really how to make sure you're prompting correctly to get all these benefits. On the needle in the haystack accuracy on that benchmark, it of course gets 100% across the board. There's some criticism of this because this is not a real use case scenario. Like usually we're not going to take a thick book with like a password hidden somewhere in that book. Then ask it to find that. Better tests are ones that kind of make sure that it's able to kind of keep context, understand directions across these large texts as it's sorting through these large texts. Looks like OpenAI came up with their own benchmark for that. That might be a little bit better. They've open sourced it. They're calling it the OpenAI MRCR, Multi-Round Coreference. So the example they gave is like they say, write a poem about tapirs and write a blog post about rocks. And they'll write multiple, like two, four, eight identical requests. So they'll be like, write about tapirs, write about tapirs, write about rocks, rocks, rocks, rocks, tapirs, tappirs, whatever, something like that, right? Then it'll say, give me the third poem about tapirs. So here the model kind of has to keep in mind how many times it was requested to write that particular poem, which one was the first one, the second one, et cetera. Reference specifically that, bring back that information about that specific poem that it wrote. And they note here that the models can be easily misled by subtle differences, such as a short story about tapirs rather than a poem or a poem about frogs instead of tapirs. What's a tapir? In case you're dying to know, this is in fact a tapir. So you learn all sorts of things on this channel, but let's continue. Here are the results for the two needle, four needle, eight needle as described before. So as you can see here, the purple line on top, that's our new GPT 4.1. So there it is. Compare that to, for example, GPT 4.0 here in green. That's this one right here. So as you can see, the accuracy is much, much better. And of course, most of the models stop here at the 128,000 context window. Whereas these new models, as you can see here, they go all the way to, you know, they have a million token context window. I did mention that before, right? So yes, they have a 1 million token context window. So matching the Gemini 2.5 Pro and some of the other Gemini models. So kind of a big deal. Definitely that seems to have a big impact on things like coding, for example, just it being able to access like just the larger context and think about it. You can kind of sense the difference in the power when you're using it, that there is something that really gets improved when you have that large context window and it's able to effectively kind of navigate through it and use it and not lose stuff, et cetera. And of course, the sort of the bigger the model, the better it is, the more accurate it is, even as you sort of increase the number of input tokens. So this is the 4.1, the 4.1 Mini, and of course the 4.1 Nano. This is the two needle one. Here is the four needle one. So definitely not quite as good. Looks like it might be better than the 4.1 at some point. And here's the eight needle. Yeah. So the Mini performs a little bit better at the very extreme and while the Nano completely drops off. So you can try it for yourself in the playground or with the API key, or if you want to try windsurf, you know, like they mentioned, they do have that free week right now. So all in all, this might not have been the most exciting release, but it does sound like the next generation of models that are going to be coming out. That's where things can get a little bit more interesting. So this is from the information and this kind of shows you the timeline of these various models getting released. Google Gemini 2.5 Pro came out in March, 2025 and currently sits as the number one across a variety of sort of user testing in LM arena, in the chatbot arena. Then the full 03 and the 04 family of models that are going to be coming out. That's really the ones that we're kind of looking forward to as the big hitters, if you will. It seems like some of the early users, according to the information, are reporting that they do believe that this will contribute to potentially discovery of new materials, new scientific information. Any number of scientists and research that we're using, even some of the previous ones, like the 03 Mini High, a lot of them have been reporting that this drastically increases their ability to conduct experiments. And this does, in fact, like maybe improve the speed at which they're able to carry out their work. And this next sort of series of models coming out could potentially take that to the next level. So I'm going to be running it through my gauntlet of tests very, very soon. Stay tuned for that. If you used the windsurf before, feel free to share any advice for somebody getting started. And for the people that are confused about why 4.1 got released after 4.5, I feel like it's pretty obvious. GPT 4.10, that comes after GPT 4.5. Just ask any AI chatbot, they'll tell you.