Opus 4.6 Fast vs Codex Spark: Fast vs Furious | The Battle for Speed

So last week we had Opus 4.6 and that six means something very specific. That six means six times the price. That's no [ __ ] Anthropic have launched Opus 4.6 fast mode which is six times the price of normal Opus but it's two and a half times faster. $30 per million um input and $150 per output. Now, the most expensive we've seen is $15 in, $75 out. So, it's double the price of what Opus used to cost for prompts exceeding 200,000 tokens. The price jumps to $60 per million and $225 per million. So, you better make you better make sure you're keeping your context down there because 200,000 tokens is um quite you you can reach it quite easily if you're a bit lazy with your your um your context usage. Uh which is crazy. We reported last week that they've given you $50 extra usage of Claude just by claiming it. I claimed fast usage. Apparently, you're supposed to just type uh /fast. So, if I go to ai.com, let's have a little look. Fast. See, it's not doing anything. You can also toggle fast mode as well by setting in the users file. So, see, I was playing around with it just before the show. We go config. Let's have a look here. Can you see fast mode anywhere there, guys? notification. No, and I have updated. So, I don't know. I'm sorry I can't bring you more news than that, but this is there. You saw by the article. You guys tell me. But at that price, here we go. $30 in uh in $150 out, $60 in, $225 out if you're over 200k tokens. And just looking here, it's also um there is a beta testing 1 million context window with Opus. This is a beta thing that you can turn on. Fast mode is compatible with that. Fast mode is best for interactive work where the response latencies matters more than cost. Now, I don't know about you, but the way we're coding, a lot of people are doing many different things all at once. And I don't know whether speed is one of those things that's important to people. Maybe you do want that space so that you can move on to something else and think about that as opposed to kind of like jumping around. And I guess that's what they're saying there when they say it's best for interactive um like working so that you can uh yeah, you are you are just needing you're you're in the flow. You're you're cooking on gas and all the rest of it. Maybe it is good for that sort of stuff. I'm not sure whether it's um whether it's worth it to be honest. But if you do go into Claude, I did claim some free usage of fast. So again, I'm unsure as to why I'm not seeing it here. But anyway, you can see the docs. We'll link the docs below where you can have a little play. Maybe it's something to do with my account. I did just downgrade to the 19 bucks plan because I want to try Codeex. I want to give Codeex a go. In a similar breath, Chad GBT also joined the party with GBT 5.3 Codeex Spark, which is now their faster model. Now, they don't necessarily have Well, I don't know. You'll probably know more about this than me, is that you can set X high, high, medium. Do you know what I mean? You can set these different modes. And so, I don't know. Let me show it to you. So, yeah, go for it. Go for it. You go for it. See, I have the the codeex and you know these are the different models. So here I'm at 5.3 codeex, but then I have the low, medium, and extra high. When was the last time you updated? Be careful. Yes. Yeah. Yeah. I've heard extra high. I updated like last night. I updated the codeex app. If you mean Okay. Okay. So this is what I and the the extra high it can take minutes minutes for a simple. So if I by mistake leave it at extra high and ask a simple ask question it might take like 3 minutes to just answer something super simple because it goes and goes. Yeah. So it's not fast at all. But I'm having fun building like four apps at the same time with it. So what plan have you got? because this uh spark okay so GPT is that GPT pro because it's a research it's a research preview on GPT pro users so yeah interesting so something interesting for our audience trying to be mindful of the costs I upgraded to the €200 uh claude code last weekend for that was I was hitting my limit constantly I couldn't do [ __ ] I was hitting the limit instantly. I was trying to do two things at the same time. It would hit hit the limit. So, what I did, I was like, "Okay, I'm going to upgrade." I did upgrade. I wasn't hitting the limits. And I immediately thought I should have listened to some and not bought the 200 model, but the the 100 one because the the 100 plan because I think that would have been enough for me. Anyway, then I switched back to codeex and I noticed I'm not hitting my limits at all. So, I just cancelled claude code. Uh the I didn't I downgraded to the 19 bucks um version. So, I still have cloud code. I still need it for some places, but the codeex with the n with the 20 euro per month is covering pretty well. It's much more cost effective. I I'm I'm like surprised I haven't hit my limits even once. I haven't been building at that level, but still I've been building. It's slower. Maybe that is also part of it. It's slowing me down, but I'm doing four apps at the same time. So, I don't know. There are pro and cons to each. Mhm. But yeah, by the I would say right now, it's kind of funny. This week, for this week, codecs have been the most cost effective for me. What I did notice is that it's running on Cerebrus and that's how they're getting the speed, which is really interesting. And I I I'm not going to claim to know much about Cerebrus, but I think these are basically I mean they're comparing themselves to Nvidia. And I think this there's mutterings at the moment and we won't get into it and I haven't done enough research on it but there are mutterings that Nvidia are gonna very quickly fall behind because there's these GPU companies like Cerebras these infrastructures. I think I think Grock with a K is one of these companies as well that have their own AI specific GPUs. Nvidia make quite generic GPUs. remember, do you remember when they released a bunch of Bitcoin mining GPUs that were very specific about Bitcoin mining? I think so. Yeah, that's good because they were specific to that, you know, you're not the Nvidia the GPUs that were previously using, they're very generic. They're trying to do lots of different things, graphics, this and that. Well, in the same way, we're going to now start to see these um AI specific GPUs. So, there we go. Very interesting. Yeah. But yeah, uh that's how they were able to achieve the the speed and maybe actually it's not a different model at all. It's actually just on a different architecture. So then saying that if we go Opus 4.6 I'm just searching off the screen now. um fast. Was there anything mentioned whether they're running it on a different architecture like how they're how they're achieving the speed? Because oh, this is just 4.6. We just want 4.6 fast mode for Claude is now in preview. Okay, this is them. Yeah. Uh yeah, I'm not too sure. But going back to 5.3 Spark, um price, they don't mention price. They don't mention anything. If you know the price, let us know down below or put on the screen. But uh presumably you either chew through your credits a lot faster or maybe it's just that. Maybe you just chew through your credits a lot faster. Maybe they're just giving it to you for free. More than a,000 tokens per second, which is insane, by the way. is insane. You'd be very happy with a few hundred, you know. But I can get this on which plan? The pro. The pro. The pro. So, I should have it. But okay. That's why I'm wondering did when did you update? Because it was it was yesterday. And in fact, I'm going to I'm going to do it my side, too, because I do have codeex. And I did find out that um even 5.3 is available. Um so it's asking me to update on my end. Yeah, I don't have it. I don't have it. Anyway, Spark, it's the it's the battle of the fast models seemingly. So, who's going to win? You decide. 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.