Start Using Opus 1m Context Window NOW

You're not using Claude at the moment, are you? I am. I'm using it for a few projects. Yes, but it's because of the usage thing. I I can't keep using it. Command AI. There's no multiplier. A a 900k token request is build at the same per. So, this to me tells me that you might as well just use the 1 million token window. This makes sense only if you need a million token, right? Yeah, but it's like why why don't you just use it? Because then you just It doesn't mean to say that you should use it like more like you should push it and push it and push it. Like if you've done your task, clear the context window, but just in case it takes a little bit longer cuz I've been doing that. I've just been doing the 1 million just in case we uh like in case I go over or whatever. Sonnet is three in, 15 out. The 1 million is three in, 15 out. So yeah, there's no there's no multiplier. Um, a nearly a nearly million token request is built at the same uh token as a 9,000 one uh 9,000 token request. So pretty cool. Um, and that includes um uh up to 600 images or PDF pages from uh 100. So quite significant. Um, and no beta header required. So I think you had to have that to gain access to it. So now here's the thing. Longer context obviously uh AIS get dumber. So whilst I encourage you to use the 1 million one just in case you go over like it's but it's but it it's not an excuse to just run it into the ground like keep keep your context as as small as possible um to maintain uh as much intelligence as possible. I'm guessing what they're saying here. A million tokens and context only matters if the model can recall the right details re and reason across them. Opus 4.6 scores 78.3 on MRCR version 2, the highest among the Frontier models at that context length. So what's the difference here? What's been going on? Input tokens. So obviously we're going all the way up to a million mean match ratio. I don't know what mean match ratio is, but is this saying Opus gets a little bit dumber? So from 91 to 78, Sonet 4.66 also gets a little bit dumber, but Sonet 4.5 actually gets a little bit more intelligent. Is that what this is saying right now? And then you've got GPT 5.4 4 who gets massively dumber over that as you increase past 256. And you've got Gemini 3.1 Pro who also gets significantly dumber. So I think they hold they hold better intelligence. But the rule of thumb here is really just don't don't if you don't need to don't go to don't push it to that 1 million. But it's nice to know if you do then it's still going to be performing quite nicely. This is part of a larger conversation on my show Command AI which we stream live every single week. We discuss the news and all things related to AI in the world of design and web. Catch us next week and join in the banter. Command AI.