Sonnet 5.5 ou Opus 5.5? O guia oficial da Anthropic

Author: Cortes do Rafael Quintanilha · 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: Sonnet 5.5 ou Opus 5.5? O guia oficial da Anthropic · Channel: Cortes do Rafael Quintanilha · 11:59 · Watch on YouTube

Topics: AI agents, Claude, Prompt and context engineering, Markets and economy

Summary

The hosts discuss Anthropic’s official guidance on choosing between Sonnet 3.5 and Opus 3.5. They describe Sonnet as a faster, cheaper fit for clearly scoped, routine, and repeatable work, while Opus is recommended for complex tasks requiring careful judgment and longer horizons. They debate whether Sonnet’s lower price makes switching worthwhile, especially for users who are not hitting usage limits, and share an example where Opus completed a more nuanced game-building task. They conclude that Sonnet can stretch a subscription’s limits and that newer models may work well with less surrounding project context.

Key points

Anthropic’s guide recommends Sonnet 3.5 for clearly defined day-to-day coding, bug fixes, and rapid feature iteration.
Sonnet 3.5 is also presented as suitable for polished documents, slides, spreadsheets, and repeatable agent tasks.
Opus 3.5 is recommended for complex work requiring careful judgment, including longer-horizon programming and knowledge work.
The hosts note that Sonnet is half the price of Opus, while shared caching can make Opus cost-effective for software development.
They argue that switching to Sonnet may not be worthwhile for users who are not approaching their usage limits.
One host describes Opus completing a Truco Mineiro game with specific rules that had not been provided explicitly.
Sonnet is suggested as a way for users on a $20 plan to make their usage allowance go further on well-scoped or repetitive tasks.
The hosts recommend reviewing project context and documentation, since long-horizon models may perform well with fewer surrounding files.

Chapters

Start	Chapter	Summary
00:00	Anthropic’s model-selection guide	The hosts introduce Anthropic’s playbook and read its recommendations for choosing between Sonnet 3.5 and Opus 3.5.
03:10	When to use Sonnet or Opus	They summarize Sonnet’s fit for scoped, routine work and Opus’s fit for complex tasks requiring more judgment.
05:14	Price, caching, and usage limits	The hosts weigh Sonnet’s lower price against Opus’s caching and debate whether users who are not hitting limits need to switch.
08:23	Long-horizon tasks and a game-building example	A host recounts Opus building a Truco Mineiro game with nuanced rules, illustrating why it may suit longer tasks.
09:56	Stretching a subscription and simplifying context	They discuss using Sonnet to conserve a subscription allowance and suggest that capable models may need less project context.

Quotes

Sonnet 3.5 comes to work on this type of task that is clearer, with a more defined scope — 03:40

If it’s more complex work, requiring more careful judgment there, including longer-horizon programming work and knowledge work, then you start with Opus 3.5 — 04:43

But I think the secret really lies in the long-horizon tasks — 08:23

The big news is that the guy with the $ 20 plan now has a model, a top model for you to use, and make your limit go further — 09:56

Sometimes with just the minimum now, like the cloud MD and a few documentation references, the model can really work well. — 11:29

Concepts and entities

Anthropic, Ed Osman, Sonnet 3.5, Opus 3.5, Claude Haiku 3.5, GPT-6, CLAUDE.md, cache, tarefas de longo horizonte, Harness

Full transcript (original language)

00:00 This is a clip from the Limite Semanal podcast that happens every Monday at 7 PM with me and Vini from AI Coders. If you want to stay on top of everything relevant happening in the world of AI, consider subscribing to the channel using the link in the description or the pinned comment. >> Let me, I wanted to bring this up, which is actually a guide, an official playbook from Anthropic, and by the way , this guy here, I don’t know if everyone knows him, Ed Osman. I’ve

00:33 known him for a long time. He’s a guy who was at Google for a long time and was recruited by Anthropic this year, okay? So he left Google, he was a big guy there, and now he’s doing a sort of DevRel role over at Anthropic. You can already, you know, you can already see the effects of Ed’s work. He is a very good guy. Well, basically, Vini, he wrote this guide here, Building with Sonnet 3.5. And there is a part here

01:04 that I think is interesting for us to bring up, which is exactly this one, choosing between Sonnet 3.5 and Opus 3. So let’s read it here together, guys. This just came out, okay? It was just published. So let’s read here together what the official recommendation is from the Anthropic team. >> Wonderful. So, in the Claude 3.5 family . Opus 3. Was built for more complex work that requires more careful

01:34 judgment, right? You will then use Sonnet 3.5. For those day-to-day tasks that are well-established, whose scope is clearer, for example, fixing a bug and quickly iterating on features, okay ? So, Sonnet 3.5. Also creates polished documents, slides, and spreadsheets, and has a good eye for design. This is interesting because I was using Opus to create some slides, man, and it

02:04 consumed a lot on Opus 3, before the update, it consumed a lot. So it pleases me if I can use Sonnet, since it’s cheaper and faster to iterate on a presentation. So that’s cool, >> interesting. >> So they continue saying that its speed, the speed of Sonnet 3.5 makes it very good for faster iterations. And look at this, your favorite, Vini, Claude Haiku 3.5 will also join the family in the

02:36 coming weeks. >> Future, view of the future, huh? Ah, that type of prediction is easy. It’s easy, right? There’s one coming >> But I saw it in the future, huh. >> You saw it here first? >> Can’t you see the screen? >> No, it’s right there, man. >> People are saying it’s not showing up. >> Wait, >> you guys aren’t seeing the screen? How come? >> Let me see. Oh crazy, it really isn’t. No, it’s not. Huh, >> damn, how did this happen? Let’s go then. Hold on, let me remove it and add it back here. >> Hang on, guys. It’s that World Cup delay. Let’s see. Actually, the idea,

03:10 Rafael, >> is that we do this so the crowd comments in the chat to see if they’re participating. Did you see that the audience went up when there was no screen? >> Because someone stopped. What are these crazy guys talking about? Just staring at nothing, right? There we go. >> Nice, guys. Sorry, for some reason on the stream it is appearing, go figure, right? >> Eh, basically I only read this paragraph here, folks. So, you who heard here didn’t miss anything, but the idea here is to say that Sonnet 3.5 comes to work on this type of task that

03:40 is clearer, with a more defined scope, right? And look, they even give a guide here. Let’s see here, okay? If your work is a programming job with a well-defined scope of day-to-day coding , that is, fixing bugs and quickly iterating on features and verifying against requirements, a very clear SDD job here, right? Then start with Sonnet 3.5. >> Wonderful. If it’s high-volume

04:10 day-to-day development 3.5 eh documents , slides, and spreadsheets, like diagrams, slide summaries, this and that. >> Sonnet 3.5. >> If it’s a well-defined task for an agent that you execute repeatedly, then investigation, review, interesting for code review, right? >> Eh, drafting, Sonnet 3.5. Now, if it’s more complex work, requiring more careful judgment there, including eh longer-horizon programming work and knowledge work, then you start with

04:43 Opus 3.5 and for your hardest problems, >> the hardest problems where you need more intelligence, Opus 3. So that is the guide, >> right, but where did it enter >> where did it enter, right? Where will Haiku enter? Oh man, >> it said >> Haiku will enter next week. Calm down >> man. I think Haiku is going to be for agent workflows, you know? Just like Luna is today. Luna 3.5. I use it a lot in agent workflows because it’s very cheap, man. For agentic workflows.

05:14 >> Yeah. Now, one thing caught my attention, Vini, which is the following , okay? Uh, Sonnet, as I said earlier in the live, is only half the price of Opus. But there’s one thing, right, that Opus is the same as Sonnet, which is precisely the cache, and that’s actually what makes Opus so cheap for software development, because software work is usually where the cache >> is most present, right? Because you’re constantly reading the same files, then running tests, running type checks, running linting, and so on and so forth.

05:45 >> That’s true. But here’s a point, Rafa, do you think they won’t improve that? No, man, because usually that value is a ratio of, uh, one to ten in relation to the output price, right? It should be like zero cents, because if they improve that, then we will see a significant change in—but do they really want to improve it? That’s the question, maybe, right? Yeah, maybe if they had improved the cache, and the cache had followed that rule of being half, we would have a bigger price difference for Sonnet 3.5. Regarding

06:15 why, the truth is, Vini, that it didn’t end up being a very big difference, especially if you want to optimize so you don’t have to keep thinking about switching models all the time. I wanted to bring up here for the guys, uh, this screen—let me confirm it’s showing now, I’m traumatized —where I basically put together the Opus 3.5. Sonnet >> in several reasonings here at the level of artificial intelligence analysis, okay? What caught my attention here? Because the big question is the following: is it worth it for me to not

06:47 use Opus 3.5 in any circumstance? My argument is this: if I’m able to use Opus 3.5 at extra high without hitting my limit, why would I try to use a cheaper one, right? If I’m not hitting my limit, maybe I can just stick with it. Oh, Rafael, but then it’s too slow, man. It’s not, its TPS is good. >> Yeah, >> so, it didn’t seem to me, it didn’t really catch my attention. Uh, unless you maybe want to do that, you know? I want to iterate over a very large volume of documents, or something like

07:18 that. I want to perform a code review on a ton of code all at once. Fine, that could work. But you know, for day-to-day, >> right? Yeah, >> yeah, for day-to-day, it didn’t seem like it to me. And so, bringing up this comparison, look at this, Opus on High, it would be tied with Sonnet on Max, but we’ve seen that Sonnet on Max pushes the cost way up, besides also being slower, right? Because it will— we know that the Max models are always much slower. So I would take Max out of the equation. Then the runner-up becomes Opus on High, okay? And then in

07:52 third, that’s where we would have Sonnet on Ex-High, which is very similar to what Opus on Medium would be , okay? >> Uh-huh. >> So, between Opus on Medium and Sonnet on Ex-High, what do you think? Maybe for the guy using Opus on Medium, would it be worth it to switch Sonnet to Ex-High? I’ll tell you… >> If you’re using Opus on Medium, if you’re using High or above, it wouldn’t be worth it in my view, under any circumstances. Unless you want to, or you are hitting your limits. If you

08:23 are hitting your limits and want to optimize them, then fine, go ahead and set up the workflow with Sonnet. If you aren’t hitting your limit, man, uh… >> But I think the secret really lies in the long-horizon tasks, >> because, you know, what I found interesting was, the day 5.5 came out, >> I had it make a game; it was actually wild because 5.5 and GPT-6 had just been released, right? And so I did a livestream, I had it make a game of “ Truco Mineiro,” okay? Truco Mineiro has some specific quirks. For example, do you play Truco, Rafa? Or not? >> Ah, back in college, right? Now, for

08:54 Truco Mineiro, you have the points there, right? Truco, you call “truco” and it’s worth four points. You call six, in Mineiro, it’s not worth six, it’s worth eight. You call nine, it’s worth twelve. So, there are some little rules like that. And so I asked it to make a Truco game. GPT-6 just made it, and I said, “Cool, nice Truco, “ but it didn’t take that rule into account. I hadn’t mentioned that rule at all. When the trick was done by Opus 3.5, man, it was something bizarre, by Zen X High. It took an hour or so to do the trick while GPT-4o only took a few minutes. But it arrived with a complete

09:26 game, like this, look. If I click to see the partner’s signal, my partner’s little character, she blinks if she has the Jack, if she has, anyway, gives a signal. So, thinking about long-horizon tasks, right, with little interaction, Opus 3.5. For me on XH, specifically on XH, it’s really worth it. Now for day-to-day tasks, its medium for me is already good enough. I’ll have to test Sonnet 3.5 a lot. To convince me of what Henrique said here, right, that Opus 3.5 in Midjourney has been very satisfying. And I wanted to draw

09:56 attention here to what Eduardo said. Eduardo Flores said that I think it is the >> maybe it’s the big news, right? The big news is that the guy with the $ 20 plan now has a model, >> a top model for you to use, right, and make your limit go further. So, if you are close to finishing yours, you are getting close to finishing your limit, you can now try to orchestrate with Sonnet for these tasks we saw. What tasks, Rafael? Well-scoped, well-defined tasks, eh, repetitive tasks, presentations, etc., etc. You go

10:27 there, throw in Sonnet, you will make your $ 20 subscription go further. Now, Vini, who could have imagined that on September 28, 2026, we would be saying that the $ 20 plan most worth it today is the one from Claude, it’s from Claude, and no longer from OpenAI, right? That >> But I don’t think even God believed in that, man. >> What a sharp change of direction. >> Eh, >> really. Eh, >> yeah,

10:57 >> you have to take your hat off. We have, we have to praise when they get it right, right, V? Of course, man. I said , even though I accessed the Anthropic accounts there in various situations, they got it right, their communication improved a lot, their public relations improved a lot. So, man, that’s it, that’s good. And you have to, you have to value it when it works out too, man. And and the model is good, man. The model, I don’t think they’re going to nerf, that they will nerf this model. You need to pay attention to the Harness, alright, guys? Pay close attention, because, you know, new models, >> especially models that handle

11:29 long-horizon tasks well, you might not need so much surrounding context files and all that anymore. And even what they brought in the CCO family, right, revisiting those files. So, it’s worth checking if your current project setup or the information inside it is making the model too verbose or if it’s actually delivering well. Sometimes with just the minimum now, like the cloud MD and a few documentation references, the model can really work

11:59 well. And honestly, man, I think we’re living in a golden age, aren’t we, Rafa ? A golden age of artificial intelligence. Mm.