🚨 Opus 4.7 is here. What’s new?

0:00All right, Anthropic just launched Opus 4.7. People on Twitter are going crazy for it. In my private chats, people are talking about it. I want to show you what it means for you, what's different. Also, Wade Foster, founder of Zapier, had early access to it. I talked with him about what it means for you. Let's talk about it. Let's break it all down here. Number one, big picture, it means it's better at coding, but you anticipated that, right? It's also got better vision, which means that it can read smaller text on on PDFs and other files that you give it. It means that it could give you better feedback on your

0:31UI. We're going to go into it more detail. It also means that it's better at instruction following. And I want to just emphasize that from everything I am seeing, many people are going to think it's worse at following their prompts because it is going to follow it much more precisely. That means if you are sloppy with your prompts before, and you were corrected [clears throat] and it understood what you said, now with 4.7, it's going to follow your sloppy prompts to sloppy outputs. And even though the prices per tokens are still the same, we are looking at a new tokenizer, which means the way that it's

1:07breaking down what you are saying as you're communicating with it is actually because it's different for some uses, it's going to take up to 35% more tokens. So, you want to be aware of it if it's if you're building it into your projects. All right, we got the benchmarks. We're going to break this out a little bit further. Big picture, here's what it means for you. It is going to require much less supervision

1:30and it's going to verify its own work. In fact, here, look, this is what Boris attorney from Claude Code has been saying on Twitter today. He's been using the word agentic a lot. Again, precise is what we talked about earlier, more precise than that. Agentic keeps coming up. Um And so, it is going to enable you to queue up tasks before going to bed and wake up to working PRs. It is going to allow for multi-steps agents with less fail. In fact, let this brings up what Wade had told me here. This is not the rumored Mythos model that is coming, the one that is supposed to be

2:07you know, really good at finding security vulnerabilities. Rather, this is Opus 4.7. So, it is the next generation in their Opus series of models. And we're our team has been doing a few benchmarks and I wanted to just share a little bit of what we're finding. So, first off, it is an improvement over 4.6. In fact, it's about a 20% improvement in accuracy for roughly the same cost. The next thing we're finding is that when you use max reasoning effort, it is scoring the highest across all of our benchmarks. So, by all appearances, this does look to be a new

2:42state-of-the-art model. The third thing we're finding is it is a very efficient model, much more than other models, specifically around multiple tool calls. And it also seems to be more precise when you have multiple steps in a workflow. Now, where is it maybe not as strong? It does tend to be fairly cautious. Particularly if you ask the model to do a really important task. You know, if you say things like this really has to get right, it will refuse to act when it's uncertain because it doesn't want to be wrong. And it also might give up early if it can't quite find all the information right

3:21away. But from everything we can tell, this is another just solid incremental improvement to the Opus models and I expect this will be a really good strong model for daily use driving for all the sort of hard tasks that you're finding. All right, let's get back into the vision, which I I mentioned earlier. We're looking at three times better vision, 3.75 megapixels. It is essentially going to be looking at your user interfaces, at your docs, at the PDFs that you send it and it's going to have better understanding of it. It's going to be able to spot bugs better and it's going to be much better at seeing. Let's keep

4:01moving on here. Again, agentic coding got better. It is 64.3% on the hardest agentic coding benchmark. That is the SWE benchmark Pro. That's the model that gets handed real GitHub issues from real open-source projects and actually has to fix them. This is incredible. It's better and I said this earlier at following instructions. It will do exactly what you say and Anthropic keeps warning, update your prompts. In fact, did Boris say this again in in one of

4:33his tweets here? He Yeah, there we go. It's going to feel much more precise than 4.6. So, again, be prepared for people to say it's not working as well, it's not as good. Um so, use simpler prompts. Say what you want more directly. And there's a new XHI effort level. It's going to sit between the high and the max effort level. This is going to um allow you to have more control over the work that

4:59it's doing. So, I like this one a lot. You're going to be able to set task budgets where you can set a cost or token ceiling at any one of the Claude runs. That means that my model is going to prioritize the work it does within your budget. And then there's the ultra reviews mode. Copy-paste that in and it's going to run a dedicated check of your work to look for bugs, to look for design issues, to look for edge cases. So again, if it's working more and more on its own, you're going to need something that will go back and review the work that it's done

5:32because you're not sitting there and doing a lot of that review as it's going on. All right. Enterprise-level knowledge, this is going to be especially useful for finance, for legal, for research. We're talking about fewer hallucinations in its analysis over numbers. It's going to be able to do multi-day research that maintains its context across sessions and it's going to allow you to automate legal research and compliance review. This is going to

5:56be a game-changer for many industries. All right, but here are a few things that you should be prepared for. Be prepared for some tasks to take up more tokens and be more expensive. Be prepared for Anthropic to just keep continuing to update, which means you might want to pin your version. If what you're building will depend on 4.7, just tell it. Connect it to 4.7 instead of allowing it to use whatever the newer

6:20and newer versions are. So, the model you're using today is just not going to be the same within 60 days. All right, that is the breakdown for today. We're going to keep looking at this and updating you and I'm going to keep doing interviews with entrepreneurs who are building using Claude Code and all the AI coding tools. Follow for more and of course, if you want this report that we've done, there'll be a link for it in the

6:42comments.