Opus Just Got Even Better!

You probably have heard the news. Anthropic have just released their biggest bad boy yet that we have access to, Claude Opus 4.7. Command AI And I I did think that there was going to be a point release based on the reinforced learning of Mythos. So, let's go through the article together. Our latest model, Claude Opus 4.7, is now generally available. Uh Opus 4.7 is a notable improvement on Opus 4.6 in advanced software engineering, cuz I actually think 4.6 took a down step from 4.5 uh in software engineering. With particular gains on the most difficult tasks. Users report being able to hand off their hardest coding work, the kind that previously needed close supervision, to Opus 4.7 with confidence. The model has substantially better vision. It can see images in greater resolution, and there was a really interesting experiment that Theo did with a puzzle. The uh was an image. You you have to a coding challenge that you have to solve by looking at an image and whatever. It's hard to visualize when we're not showing it. However, he handed the image, and the image was um it it it got flagged as dangerous, well not dangerous, that's the wrong word, but it got flagged basically as Opus 4.7 with with its increased security capabilities has flagged this as, you know, conspicuous or whatever. You saw it. Kind of interesting. So, yeah, high resolution in combination with its um yeah, it it's it's focus on cybersecurity is is causing a bit of whatever. So, looking at the benchmarks, again, we don't care too much about benchmarks, but I think it's it's nice to see a comparison cuz we've got GPT 5.4 here, Gemini 3.1 Pro, and then of course the big daddy Mythos. So, um is and and I obviously Opus 4.6. So, big improvement on agentic coding. Uh big improvement on the sweet bench verified on agentic coding as well. And a you know, I think Well, no, it is an improvement cuz again, this was terminal bench actually fell in Opus 4.6, but I think we've gained some ground here. Uh but of course, GPT 5.4 beating it anyway on terminal bench. So, there's that. Uh Humanity's Last Exam, increasing there. Uh 5.4 was was better than 4.6 on the on the uh Humanity's Last Exam anyway, but overtaken again. Just to bring in mythos as well. Still not beating mythos on any of these, to be honest. So, that's crazy. 90 93.9 anyway. Um so, yeah. As you can see, we won't go through all of these, but it's uh it's an improvement. But I think the biggest thing with regards to this model is just the way it handles things. Now, the I'll find it. Below are some highlights and notes from early testing Opus 4.7. Instruction following. Opus 4.7 is substantially better at following instruction. Interestingly, that means uh that this means the prompts written for early models can sometimes now produce unexpected results. What this is telling me is that you guys need to actually um adjust the way that you work because of this model. Now, I I would say in the last few weeks, I've started to really adjust the way I work. I used to be very prescriptive on the exact files it needs and this and that. Whereas I've now need I've now more lax with my prompting because I know that it does a really good job at figuring out what it is that I want and what files it needs to access. This might mean that you probably need to spend some time with it to figure out how you how it works for you and the way you like to work. So, uh previous models interpreted instructions loosely or skipping parts entirely. Opus 4.7 takes the instructions literally. So, maybe we need to probably be a bit more prescriptive. A bridge between what I was doing before, which has been very prescriptive about the files and this and that. And and the lax approach that I've taken recently, probably just need to hone in a little bit more cuz it will literally take literally. Improved mode of support. 4.7 has better vision for high resolution. We've already covered that. Real-world work has a well state-of-the-art score on the finance agent evaluation. I really want to build a finance agent. I think that's something I really want to do and use Opus for it. So, yeah. What to do to access your bank? No, to um trading. Trading. Stocks and shares. To do like actually give it a hundred bucks and then by the end of the day it returns a thousand. Yeah, let her rip. Um hey mate, if Fin Bench if Fin Bench is anything to go by, I wonder if they do talk about it in here. Memory 4.7 is better at using file system based memory. So, I'm thinking if you're using uh Obsidian all of that stuff would be good for Obsidian users. Uh knowledge work, which is GDB val um yeah, kicking butt. I wonder what um Mythos would be in this thing. Safety and alignment. Overall, Opus 4.7 shows similar safety profile to Opus 4.6. Our valuation show low rates of concerning behavior such as deception, syco- sycophancy. I hate that word. And cooperation with misuse. Some measures uh such as honesty and resistance to malicious prompt injection attacks. So, that's quite interesting as well. Alignment conclude the model was largely well aligned and trustworthy, though not fully ideal in its behavior. That's quite concerning, don't you think? So, misaligned behavior um I I mean, is is lower better? Cuz I think we've regressed here, boys and girls. I think lower is better. Misaligned behavior score from our automated behavior unit on this evaluation, 4.7 is a modest improvement on 4.6 and 4. So, it is higher is better. So, what? Mythos is misaligned. Anyway, so there we go. And just in amongst all that, which is your favorite mode? Max. So, this is extra thinking. Now, this also causes the idea, you know, ultra review as well as something I saw, which is I saw in the Yeah. Um Uh produces dedicated review session that reads through changes and flags bugs. So, it's basically coming after Code Rabbit. Opus 4.7 is a direct upgrade to Opus 4.6, but two changes are worth planning for because they affect token usage. Now, this is the big one. Opus 4.7 uses an updated tokenizer that improves how model processes text. The trade-off is that the same input can map to more tokens, roughly 1 to 1.35 times, depending on the content type. Second, Opus 4.7 thinks more at high effort levels. So, you might want to rethink that max mode cuz it gets exponential. Particularly on later turns in agentic settings. This improves its reliability on hard problems, but does mean it produces more output tokens. Users can control token usage in a variety of ways uh by effort parameter, which we just spoke about before. So, yeah. I mean, look at this. Basically, it's exponential. It jumps the same amount of tokens at as it took the same amount of tokens as it took from low, medium, high, extra high, it takes to get to max. That's that's a similar. Um but yeah, if you have harm problems, you might want to utilize it. I don't think you need max, mate. I really don't. Experiment with extra high. See how far you get with extra high. You won't, will you? Uh Can't bring yourself to it. That max button is there for a reason. Mhm, mate, we're already struggling here. We're already struggling for GPUs and there's you max mode all the time. Max mode rendezvous. 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.