# Anthropic Just Dropped the Feature That Makes Sonnet Feel Like Opus ## Метаданные - **Канал:** Ray Amjad - **YouTube:** https://www.youtube.com/watch?v=rj7xi-s4Ssg - **Дата:** 09.04.2026 - **Длительность:** 5:23 - **Просмотры:** 16,930 - **Источник:** https://ekstraktznaniy.ru/video/48856 ## Описание Level up with my Claude Code Masterclass 👉 https://www.masterclaudecode.com/?utm... Use LASTCALL for 35% off lifetime (before lifetime plan is retired on Friday 24 April 23:59 PST) My Claude Code newsletter 👉 https://www.masterclaudecode.com/news... My Claude Code coaching 👉 https://www.masterclaudecode.com/cons... I've never accepted a sponsor; my videos are made possible by my own products... —— MY CLASSES —— 🚀 Claude Code Masterclass: https://www.masterclaudecode.com/?utm... —— MY APPS —— 🎙️ HyperWhisper, write 5x faster with your voice on MacOS & Windows: https://www.hyperwhisper.com/?utm_sou... Use coupon code YTSAVE for 20% off 💬 AgentStack, AI agents for customer support and sales: https://www.agentstack.build/?utm_sou... 📲 Tensor AI: Never Miss the AI News on iOS: https://apps.apple.com/us/app/ai-news... on Android: https://play.google.com/store/apps/de... 100% FREE 📹 VidTempla, Manage YouTube Descriptions at Scale: http://vidtempla.com/?utm_source=yout... ————— CON ## Транскрипт ### Intro [] Okay, so we have a brand new slash command added to Claude Code, which is likely preparing us for when Claude Myphos releases. So it's /advisor. So when you do /advisor, you can configure the advisor tool and it should take you to something like this. And essentially this is a tool that's built into Claude Code for when it requires stronger judgment about a particular problem. Now you can see right now we only have Sonnet and Opus. So if you're on Opus, ideally you want your main model in Claude Code to be on Sonnet. So you do /model and then pick Sonnet. And essentially when you're coding with the Sonnet model and it ends up getting stuck, then it would call the advisor tool to ask Opus a question. This is kind of similar to /model opusplan whereby Sonnet is used all the time except for when you're in planning mode, in which case Opus is used. So if you were to combine the two together, you would have a really good plan at the very beginning ### Demo [0:53] then it would switch over to Sonnet for the actual execution. And if Sonnet ends up getting stuck with something, then it would call the advisor tool to ask Opus what the deal is here. Now right now with /advisor, you can only see that we have Opus 4. 6 and Sonnet 4. 6. So if you're already on Opus and you're using Opus as advisor, you'd probably not see much benefit to doing that over using a subagent of sorts. And if you're in Haiku, then you may want to use Sonnet 4. 6. Now what they will likely be adding here in the future is like Myphos for their upcoming model, because chances are since it's so expensive, we won't be able to afford it most of the time. Until like prices come down and stuff in the future. So I imagine what many people will be doing in Claude Code is sticking to Opus and then having the advisor call Myphos with the entire like chat history and log to help it work around like a particularly hard problem. Now the way that it works behind the scenes is that essentially the tool description consults a stronger reviewer who sees your full conversation transcript. There are no parameters. When you call advisor(), your ### Comparison [1:52] entire history — task, every tool call and result, your reasoning — is automatically forwarded. The advisor sees exactly what you've done. Now, of course, you can tell Claude Code to call this tool when you're stuck, but it says it will be called before any substantive work. So before writing, before committing to an interpretation, before building an assumption. And it's also called when the model believes the task is complete. So it's told to first make deliverable durable first and then actually call the advisor. When stuck, errors recurring, approaches not converging, results don't fit, or when considering a change of approach. Now guidance on frequency says, on tasks longer than a few steps, call it at least once before committing to an approach, and once declaring done. On short reactive tasks, skip it. Which means that if you ask Claude Code to change some colors around, then it will not call the advisor because it's too simple of a task. Now this is a part that's interesting. So how to weigh the advice. Give it serious weight, as you would expect. And if the empirical results contradict advisor advice, surface the conflict with another advisor call rather than silently switching. So I find this part particularly interesting. I'd like to see that in action. And now it's like a passing self-test is not evidence that the advice is wrong. Now this is a summary of the tool description. If you want to read the full one, then it will be down below in the pinned comment. Now let's see this in action. If I go to previous session and this is some bug that Claude Code found with the billing logic ### How It Works [3:11] if I wanted to call the advisor, I could either like tell it directly or Claude Code during the execution may realize that it needs to call the advisor. So we already have it set over here with Opus. And I will just say, hey, can you call the advisor to check that this will be the solution here? And then pressing enter, it now says advising using Opus 4. 6. And the advisor basically flagged two things over here that we would want our Sonnet model to consider. Now it seems that there are some things to bear in mind when doing this. Firstly, it seems that the advisor like can't actually consult extra files on your machine. The entire chat history is being passed to the advisor. And then that result is being passed back into your session. So as they showed on Twitter, this is what it looks like. You have the main executor being Sonnet. You have the shared context that has been accumulated so far in the history, which is a conversation, tools, and so forth. And then the executor can call the advisor on demand. That advisor would send feedback and that would be added to shared context, which likely means ### Comparison to Slash Loop [4:08] that any future advisor can see any previous advice that previous advisors have given. Now, when it comes to benchmarking, they found that using this approach is slightly cheaper if you were paying for raw API tokens, but it would mean that it consumes less usage limits on your Claude Code plan as well. And you also do get slightly better performance on some of these benchmarks. Now I will personally not be using this that much because I use Opus all the time and I don't see any benefit in getting another Opus advisor with the same chat history so far to give advice. I'd ### Stream vs Poll Commands [4:37] rather get advice from a subagent that has an independent context. But I would see myself using this in the future whereby if Myphos is released, basically swapping this all around, I would have the main executor being Opus because it's still a very capable model. And then getting advice from Myphos instead. And this is probably one of the ways they're planning on making Myphos accessible to us once it is released, because apparently it will be really expensive. Now, if you do like this kind of stuff, then do subscribe to the channel because I do make the most comprehensive Claude Code videos here on YouTube. And if you want to stay ahead of the curve, then I also have a Claude Code newsletter whereby I share some of my own strategies and techniques ### Examples [5:11] like on a regular basis. Signing up will give you access to some free videos in my masterclass that you may find helpful as well. There will be a link down below if you are interested.