12 Segredos do Opus 5.5 Revelados Pela Anthropic (Mais Rápido e Barato)

Author: Felipe Borges - Fala IA! · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: 12 Segredos do Opus 5.5 Revelados Pela Anthropic (Mais Rápido e Barato) · Channel: Felipe Borges - Fala IA! · 17:01 · Watch on YouTube

Topics: Claude Code, Prompt and context engineering, Software engineering with AI, AI content generation, AI models

Summary

The video presents 12 practical tips for getting better results from Anthropic’s Opus 5.5 in Claude Code, with a focus on balancing quality, speed, and cost. It recommends starting with medium effort, then testing effort levels against real tasks, and explains how to preserve context when changing effort within the same model. Other suggestions include using an agents.md file, closing completed answers, tracking long tasks with a to-do list, and signaling when time matters. The video also advises naming design patterns and providing tools such as Python or OpenCV to help the model inspect specific parts of images and charts.

Key points

Start with medium effort as a likely cost-benefit sweet spot, then test low, medium, and high effort on tasks from your own workflow.
Higher effort does not always produce better results, especially for simple tasks.
Use agents.md for shared model instructions; the video says Opus 5.5 can read it when cloud.md is absent.
Within Opus 5.5, effort can be changed without rereading the context cache, but changing models mid-task can require rereading it.
The video mentions a temporary usage-limit reset, available only until October 22.
The model no longer displays its reasoning, according to the video.
Tell the model to treat an answer as closed unless a later message asks about it or identifies a problem.
For long tasks, maintain a task list so the model can track remaining work and avoid declaring completion too early.
Provide a time budget or state that time matters when you want faster execution.
For stronger design and image-analysis results, provide design references and systems, name design patterns, and offer tools such as Python or OpenCV.

Chapters

Start	Chapter	Summary
00:01	Opus 5.5 and effort levels	The speaker introduces the guide and explains why medium effort is presented as a strong starting point for balancing results and cost.
03:04	Test effort and standardize instructions	The video recommends comparing effort levels on real tasks and describes using agents.md alongside cloud.md.
06:10	Changing effort and usage limits	The speaker explains changing effort while preserving cached context, describes a temporary limit reset, and notes that reasoning is no longer shown.
09:17	Keep answers closed and track long tasks	The video suggests preventing unnecessary revisits to prior answers and using a task list to keep long jobs on track.
11:20	Time budgets and design guidance	The speaker discusses setting time expectations and explains why design references and named patterns help guide the model.
13:54	Tools for image analysis and recap	The video recommends giving the model tools to inspect specific image regions, then summarizes the tips and closes with a call to action.

Quotes

I choose a real task that I do every week in this project — 03:35

treat that answer as closed. — 09:47

telling him that time matters. — 11:50

you have to provide good references and know how to ask it for them — 13:24

Concepts and entities

Opus 5.5, Claude Code, Anthropic, AGENTS.md, CLAUDE.md, níveis de esforço, cache, Python, OpenCV

Full transcript (original language)

00:01 If you’re using the new OP 5 Cloud Code , be careful, because many things that worked in previous models are now more expensive or significantly slower in your workflow. I didn’t say this; Antropic itself published a complete official guide on how to get the most out of this model. And of course, I thoroughly analyzed the documentation and selected 12 practical tips that will make your development much faster and, most importantly, much cheaper. So , if this interests you, grab your coffee and let’s get to the video. Everyone already knows that the OP 5 Cloud Code is a very powerful model.

00:31 It’s up to 40%cheaper than the OP 5. It’s also faster in output, meaning in the responses it gives you. It has up to 60%lower cash readout, and I’ll talk more about that later, but what really matters is how we can get the most out of this model according to the documentation. And I’ll start right here with rule number one, which is the main one for you to start using OP 5.5 at medium effort. And for those of you who don’t quite know what the effort of a model is, if we look down here, within our Cloud Code, we see that we

01:01 can select the model. It’s selected here as OP 5.5, and the effort, which is how much this model will reason to deliver the task you asked it to. And notice that we have several levels of effort. We can go from low to ultra code. And the further to the right, the smarter, but more expensive. The further to the left, the less smart, but faster and cheaper. In the past, OP 5 had the best of both worlds, the best cost-benefit ratio, precisely the high effort. So this was the effort where OP 5 delivered the best result for the

01:32 lowest price. Now that has changed, and we’re now talking about medium effort. And to make this very clear, take a look at this graph that will show precisely this cost-benefit relationship. Here, at the low effort level, it will deliver a result of 87.4 %and cost 11 cents. But when we take that to the medium level, look at the jump it makes. It goes to 92.8 and increases the cost by 0.21 cents. If we continue going up, it improves even more. Of course, the model improves,

02:02 you’re increasing its effort, but then the cost rises to 29 cents, it goes to 74 cents for an improvement that isn’t that big. If we look at this other benchmark, it also shows the same thing . In this case, the cost to perform the tasks related to this benchmark was 84 and it achieved a score of 86. To achieve a slightly higher score of 88.4 , it already spent 1.38. Now, we know that benchmarks aren’t the final word. We always have to look at benchmarks with a grain of salt, because they

02:33 don’t necessarily represent our day-to-day work. But when it comes to effort and its cost, it’s a strong indicator that using medium effort is the best option. And don’t worry, I’ll leave all this documentation referenced in the description. If you want to read more, understand more, it will all be in the video description or in the first pinned comment. But moving on to our second rule for using the model well, what I recommend is: always test the effort levels according to your work. Benchmarks are good, they give us a good idea, especially regarding cost and benefit, but only you know what you

03:04 do in your day-to-day work. Eventually, you will need to use high, extra-high effort for certain tasks, and not for others. Everyone has a different workflow. So test to identify what is the best cost-benefit in your case. In general, it will be medium, as we just saw, but depending on the task, you may need more or less effort. I myself tested it in my workflow and I identified tasks that I do need to use extra-high or some that I’m using medium. And there are even tasks that I use on low effort. And to help you

03:35 discover this, I’ll show you the prompt I set up myself to run this test. Here’s a simple prompt, an effort test for us to do, and it says: “I choose a real task that I do every week in this project to ask it to look at the file history and run that same task three times in three sessions: one on low effort, one on medium, one on high.” This will give you a completely personalized comparison level according to your tasks. So, run this because you’ll be surprised. After all, a model with higher effort doesn’t necessarily mean better results. This is because the more your model has to think, if

04:05 the task is very simple, it ends up bringing a lot of complexity to it and it can deliver an even lower result than if you were using the correct effort. But now, here in rule three is one of the coolest things. Finally, Antropic decided to do this and standardize our PD document. When we were using the other models, this also applies to the previous models. We had to include an instruction document for our model called cloud.MD. In this document, you put instructions on how your work should be done, what you wanted, and what you didn’t want in the results. And it had to have this format

04:36 : cloud.md. Therefore, it only worked in the cloud. This document, just so you understand, is a document that the model automatically reads whenever it enters the workbook. So, if this document didn’t have the name cloud.md or FMD, the model wouldn’t read it automatically. But the problem with this is that all other models, for example, when you’re using the Codex, will read a document called agents.md. And it’s exactly the same document idea , the same structure, just with different names. And now Antropic has placed agents.md inside our cloud code. In other words, the model will now enter your folder, search for the

05:07 cloud.md model, and if it doesn’t find it, it will automatically read widgets.MD. So this becomes much cooler , much more standardized, because now you don’t need to have multiple documents and update several different documents, depending on the model you have to use. And if everything I just told you sounds like Greek to you, you need to check out our Wayi Verso community. There, you’ll learn everything you need about Cloud Code, how to use AI in the best possible way, and you’ll have access to my black box, my creation process, so you can build websites like these, images like the one you’re seeing, and also have access to my Premiere Hey tool, so you can do

05:39 professional video editing with just a few clicks of the mouse. In addition, you’ll also have access to the entire AICloud track, so you can learn everything you need about Cloud Code, how to work with it professionally, as well as tracks like Google Gravy, N8N, and all the content and courses I’ll be adding. From now on, you’ll have access to all of this without any additional cost. But in my opinion, even cooler than everything I just told you is the black box, which is where I share everything: skills, tools, all the knowledge I use to develop all the material I create here for the channel and also for my clients. Therefore, if

06:10 working with Cloud Code professionally interests you and you want access to our community, ask me questions, access exclusive materials, scan your phone at the QC code that’s appearing on the screen and come be a part of Everse. You’ll be very welcome. But continuing here, the next new feature is also very cool, because now we can exchange effort without losing cash. So imagine you’re here in your Cloud Code asking it to do a task that, for example, involves the Max plan, right, the Max effort within OP 5. And eventually it finishes doing this task. But then you want to ask for a simpler task. For example, execute this code so I can find it in the browser. So you think: “

06:42 Damn, I don’t want to use OP 5. In the Max version to perform such a simple task.” What can you do now only in the Opus model? You can change the effort. So, imagine you’re in Max, you’re going to ask for a simpler task, but still within the same context. You come here, set the effort low, and you can ask it to execute or change something in your CRM or on your website that are simpler tasks. “Hey, Felipe, but I could always do that.” Yes, but now this is worthwhile. Previously, what we could do was change the effort, as well as

07:12 change the model. But whenever we did that, changed the model or the effort, there was something called memory cache . Everything we discussed, all that context that the model had to understand to perform its task, everything you discussed was saved in memory. When we changed the effort, the other model with less effort, or the other model we simply chose, had to read all that cache again. Guess what? It used to take a lot of time and tokens, so it cost money to do this. You thought you were saving money, reducing the model’s effort, for example, to the lowest possible level,

07:43 but in reality you were spending more, because changing the model would require rereading all that cash again. That cost money. Now, within OPO 5, you can change the effort here without losing that cash memory. Therefore, the model won’t have to read it again. That’s great. But remember and pay attention, it’s the effort you can change. You can’t change the model mid-task, because otherwise the same problem will happen to you. You’ll lose the cash and it will have to read everything again. And what would be a saving, you’re actually spending much more. New feature number five is that

08:15 we now have a limit reset. So when we go into our cloud code, you’ll see that it says “redeem a bonus credit.” And if you also go to “usage,” you’ll see that we get a free reset. Gain more leeway to explore OP 5.5. Imagine you’re exceeding your weekly limit or even your 5-hour limit and you want to reset , go back to zero. You can, you’ve now earned this right with the OP 5.5 model , but pay attention because this is only valid until October 22nd, okay? So if you’re watching this video later, everything applies to you except this

08:45 part. If you’re watching before, use it wisely, wait to reset until your limit has been reached and you really have a use for it. Otherwise, keep this ace up your sleeve, it’s very useful. New feature number six, and this one isn’t so good, is that now we can no longer ask to see the reasoning. Imagine you asked it to perform a task, it performed that task, but you want to know how it performed it, what the reasoning behind it was. Now this will no longer appear to you, precisely because Antropic doesn’t want competitors to have access to how the model reasons. So, if you previously had a use for this, well, know that in these new models you will no longer

09:17 have access to that kind of thing. However, honestly, I don’t think it’s really missed, and Antropic is right to want to protect how its model works. The seventh new feature is that now, if you’re interacting with this model, it will keep revisiting other answers it has already given you. This can be very annoying when you want to have a quick back-and-forth, ask a few questions that it wouldn’t need to consult. So, to solve this problem that will happen to you and to also spend fewer tokens on this, I advise you to add a prompt ( either written out, in the message you send to the model, in cloudmd, or even

09:47 in agents.m, which it now reads) that says the following: “After you have answered something, that is, after it has given you a first answer, treat that answer as closed.” In future messages, focus your reasoning on what I’m asking and don’t go back to a previous answer unless I ask about it or point out a problem in the question. By sending this to him, he will stop revisiting all previous answers unnecessarily, unless there is a need or you expressly ask him to. This will greatly help your use of Appus 5.5. New feature number eight, which is a rule of use that will help you a lot, is

10:18 that you maintain a to-do list. Appus 5. It’s constantly revisiting its work. So, it’s there making a to-do list and it will keep asking if something is ready. Eventually, one of the items on the task may be ready, but not the task as a whole. And eventually, it may say that the task is completed when it is not. To solve this, it’s very simple, just keep a file called task.mmd. So, whenever you are going to execute a long task, and yes, I’m talking about long tasks, simple tasks, this won’t happen. Then you send this prompt here, keep a list of Tasks within tasks. MD

10:49 with one line per endpoint and per test . So then it will have a task list where it can look at this list and check it off, even if the model keeps returning, look, this checkpoint is ready, look, this item is ready, this is ready. It will look at the task list , see what’s ready, but it won’t forget what still needs to be done, because that’s what it does. It sees that something is ready, forgets what it has to do and ends the task there. Now, when you have a task list MD, it will look at the list. As long as all are not complete, it will not mark them as completed. This is very important. This

11:20 prompt solves this and this was taken from the documentation itself, as written here. New feature number nine is that now you can pass a time budget to the model. What does this mean? Imagine that you are taking too long and you want it to perform a task as quickly as possible. You want, I don’t know, it to perform something in 3 minutes. You can pass this to it. Look, I want everything done in 3 minutes. So he’ll organize his execution to deliver before the deadline you gave him. This can be done so that he delivers a task a little faster. But honestly, this time budget thing isn’t my favorite way

11:50 to use it. I tested it, and I didn’t like it. This, however, I found really good: telling him that time matters. Imagine you want a task and you want him to execute it quickly, but you don’t want to compromise on quality or items you want him to complete in the task. So you can simply say that time is important to you. What he will do is prioritize time over quality. Not that the quality will be bad, in fact it wasn’t. The results I got were very good. What I saw was a result practically as good as if I hadn’t asked for this. But with a considerable difference in time. This isn’t my invention or hallucination. It’s in the

12:21 documentation. If you can’t predict a time budget there, you can tell him that time matters and yes, the The model will work a little faster. Here are some of the results. Again, everything I’m saying is in the documentation. You can look at it and delve even deeper. And to understand how to apply this, it’s very simple. So you’ll send, for example, a list of the clinic’s direct competitors in the region, compare prices and services, and give me a table with what we do differently. Time matters here. Don’t waste it; you can avoid it. And the sooner you get a correct result, the better. That’s it. It’s that simple. It will deliver the task to you in a much

12:52 more efficient way. New feature number 11 that I found in the documentation, and it concerns one of the things I like most, which is design. It clearly tells you to name the design patterns. That is, in Portuguese, of course, here , if you use OPO 5. Good references, which is what I always talk about, and I talk about this a lot in the Aiverso community, it will deliver generic results. Yes, the OPO 5 model. It’s the best model I, Felipe, have ever used. To create designs, but if you don’t provide good references and your own

13:24 design system, it will deliver the infamous AI Loop. And the detail is, the documentation itself says that if you ask the model to avoid that pattern generated by AI, what it will do is switch from one pattern to another. That is, imagine it returned everything in purple and magenta palettes, it will now switch to a black, red, and white palette. These are patterns, and what it will do is just change those patterns. Therefore, I repeat, if you want beautiful design, you have to provide good references and know how to ask it for them, know how to create a design system. A good model won’t necessarily deliver a different design

13:54 if you don’t know how to guide it to that. And again, if you want to know how to do this, the community is the best place, mainly because I really enjoy this design aspect. And there you will find everything you need, an operating system for design, my skills, in short, it’s the right place for you if you want to learn more about this. And now, drumroll please, because the 12th new feature is that you can give it tools to…So that he can analyze graphs, images, and diagrams in a much better way. And how are you going to do that? By giving him Opus 5. This model is already very good at reading

14:25 diagrams and images, but you can make it even more efficient. For example, imagine he needs to see this graph you’re showing and you want him to focus on these specific months. Before, you would have to show a printout of the entire image and explain through prompts where he should focus and what he should change. Or if he needed to cut it out to analyze only those data points, for example, he wouldn’t be able to. Now, if you provide him with tools that allow him to do this, such as Python or OpenCV, which are Python tools. If you simply give him that, he

14:59 is then able to focus his attention on that section, on those specific data points where you indicate he should focus. Now, finishing up the 12 rules in a single glance. Start with the medium. This will be the best cost-benefit ratio for most tasks you will use it for. But, obviously, we can’t simply accept that everything will work. The same for everyone. Therefore, test your workload levels. Another cool thing, use agents.md. Okay , cloud.md still works, but eventually if you want to use another model that only reads agents.md, it would be very annoying to have to duplicate and

15:29 change it. So now just create agents.md as the default, which will be much better. I even think that cloud.md MD will eventually become obsolete with this. Another thing, whenever you are doing a task, pay attention to the level of effort you need to use. If you needed a higher effort to create one part of that task, and the other is a simpler task, a simpler process, you can now change the effort of the model. As long as everything is within OPU 5.5 , you won’t have any problems: it won’t have to read the cache again, the cache will be preserved. Then, keep your limit reset ace up your sleeve. Don’t

16:00 ask to see the reasoning because he won’t show you anyway. Answer given is answer closed. You will have to pass. Remember, this is for him? So he doesn’t have to keep reviewing the answers he’s already given you. Keep a to-do list to ensure that he delivers everything correctly on long tasks. Give him a time budget, if you have one . Or, what I prefer, say that time matters. Then, name the design patterns , because a good model doesn’t necessarily deliver good design, unless you know what you’re doing. And give him a magnifying glass of where he needs to look, if you’re developing a chart, something, a diagram; this will greatly improve your results. But now I

16:31 want to know if this video helped you, if you had already seen any of these tips, if you’re already applying them, and which ones you’re going to apply. So tell me, leave a comment about what you thought and what you’re going to put into practice today. Furthermore, I remind you again that if you want to work with Cloud Code professionally and have access to all my tools and processes, you need to be part of Aiverso. Just scan the QR code on the screen with your phone, come join, and you will. Having access to everything that’s already there, and there’s a lot , but also everything I release, whether it’s tools, lessons, or exclusive material. In addition, I also ask that you subscribe to the channel,

17:01 leave a like, and help spread the word so this video continues to reach more people and we can grow our community. Now I’ll leave you with another video so you can continue learning about artificial intelligence. And I’ll say goodbye now. A big hug and bye. Yeah.