YT YouTube2Text

YouTube2Text → Transcripts → Claude Fable 5.2: Anthropic Says Use Opus for Most Work

Claude Fable 5.2: Anthropic Says Use Opus for Most Work — Transcript

by Dylan Davis · 2,807 words · 375 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Claude Fable 5.2 just came out this week
0:03and it's the smartest model that you can
0:05get your hands on, at least right now.
0:07But here's the surprising part. The
0:08company that built that model is telling
0:10you to start with a different one for
0:12most of your work. And I actually agree
0:14with them. I work with my clients every
0:16single day, and Opus 55 handles 80 to
0:1990% of their work perfectly. Fable just
0:23burns through your plan a lot faster.
0:25So, let me show you how I choose between
0:27the models and the only two times I'd
0:29actually bring Fable in. So, let's get
0:31into it. So, I wanted to run a little
0:34test just to prove to you that Opus 55
0:36versus Fable 5.2, the quality is the
0:38same for most knowledge work. So, in
0:40this test, I took 15 files, all of which
0:42are pretty decently sized. So, some were
0:43PDFs or above 50 pages and some were
0:45Excel files that were quite big and had
0:47many rows. So, I had 15 files and there
0:50were nine hidden problems in those 15
0:51files. I then gave both of these to both
0:54Fable 5.2 and Opus 55. And I simply give
0:57it a request asking, "I'm about to buy
0:59this business. Read everything and tell
1:00me what I should worry about before I
1:02sign." Now, all this data is fake. It's
1:04synthetic, all that stuff. But I wanted
1:06to just show you the results. So, when I
1:07ran these two side by side, Opus 55
1:10Medium and Fable 5.2 Medium effort 2,
1:12they both came out with the same exact
1:14nine issues. Fable 5.2 did come up with
1:17an extra point, but actually didn't
1:18change the decision at all. And that was
1:20just one of the many tests that I ran
1:21because I wanted to show you the
1:22similarity in the quality between the
1:24two models. The other test that I ran is
1:26that I had I think around 65 invoices
1:28and I gave them both to Opus 55 and
1:30Fable 5.2 and I wanted to match those
1:32invoices up to bank statements. Again,
1:34they both did it perfectly and similar
1:36time windows as well. So this is just a
1:38small sample test that I ran here, but
1:40this is from my daily experience. Again,
1:41I work with people every single day and
1:43I see them using these models time and
1:45time again and the quality they get from
1:46Opus 55 is almost always the same as
1:48Fable 52 irrelevant of the task,
1:50relevant to the complexity of the task.
1:52So now before we get into the details
1:53around Opus versus Fable and the
1:55different effort levels, one thing I
1:56want to flag is when I work with people,
1:58the first thing I want to do is have
1:59them share their screen so I can see
2:00what model choice they have because they
2:02might come to me and say, "Hey, you
2:03know, Claude sucks for this specific use
2:04case. I kind of hate it." And as soon as
2:06they show me their screen, I can quickly
2:08see exactly why that's the case. Often
2:10times you're using the wrong model at
2:11the wrong effort level. You might ask
2:13yourself where is this? Well, here I
2:14have the claw desktop app and you can
2:16see here at the bottom right hand side
2:18it says fable 5.2 high. When I click
2:20this you can see we have the different
2:21models that we can choose from. We have
2:23fable, opus, sauna and haiku. Everybody
2:25I work with we usually stick with sonnet
2:26opus and fable. Haiku is kind of
2:28irrelevant for most activities. And then
2:29below that we have the effort level. So
2:31this basically highlights how hard the
2:32AI is going to try any given task. So we
2:35have between low and max. We'll talk
2:36about all these in more detail, but I
2:38just wanted to highlight where that sits
2:40and how you can change it. And quickly
2:41note, Haiku is the smallest one, sonnet
2:43is medium, opus is big, and Fable's the
2:45biggest. Quick pause. This video is
2:48brought to you by me as always.
2:50Something I hear a lot when I first talk
2:52with people is they say something like,
2:54"We're paying for Chatbt or Claude or
2:56fill-in- thelank other AI provider and
2:58we're barely using it." Or, "I tried
3:00building something myself, got stuck,
3:02and then I kind of just gave up, went
3:04back to the manual process." If that
3:06sounds familiar, that's what I help
3:08with. I work hands-on with leaders and
3:10their teams. You share your screen and
3:13we build an AI tool around your actual
3:15work and then you leave the session with
3:17that tool working. If you want to see if
3:20there's a good fit between the two of
3:21us, there's a link below. You can grab a
3:24free call as well as see the offering.
3:26And with that being said, let's get back
3:27in the video. Now, when people share
3:30their screens with me, there are two
3:31very common mistakes that I see. So one
3:33is they come to me and say the quality
3:34of the output from the AI is really poor
3:36and I'm not happy with collab for
3:37whatever reason. They show me their
3:38screen. They're likely using a small
3:40model at a lower reasoning effort. For
3:42most people that's usually sonnet 55 at
3:44medium effort because that's what
3:45sometimes your enterprise your IT team
3:47sets your default to. Any conversation
3:49you start it's always going to be at
3:50sonet 55 unless you change it. That's
3:52the first issue. So the model's too
3:53small. The second issue is that often
3:55times people will think that bigger is
3:57better. So I'm going to use the biggest
3:58model at max effort to do many of my
4:00activities. This is a great way to burn
4:02a ton of money on accident or
4:04intentionally thinking that you're
4:05getting better outputs. So, if you have
4:07a set allowance for your seat in your
4:08company, you're likely going to burn
4:10through that allowance of either money
4:11or subscription percentages within a few
4:13hours if you're using Fable 5.2 at max,
4:15maybe even a few minutes depending on
4:17how much allowance you have and what the
4:18task is. The best situation is that
4:20you're likely starting with Opus 55
4:22medium. That's what I recommend all my
4:24clients start with. You start with Opus
4:2555 at medium effort. That covers most
4:28knowledge work. And when I say most,
4:30probably 80 to 90% of the tasks that I
4:32do with my clients can be done with Opus
4:3455 medium or high. And when I say done,
4:36done very well. And I'm not alone in
4:38this. Anthropic, the company that made
4:40these models, stated specifically in
4:41their own documentation saying that Opus
4:4355 at medium and high effort is the best
4:46daily driver for knowledge workers. And
4:47to that point, there is a huge gap in
4:49cost between Opus 55 and Fable 52. Some
4:52people think that Fable 52 might be just
4:53slightly more expensive than Opus. It's
4:55much more expensive than Opus. So
4:57there's tons of benchmarks out there
4:58where people test these different
4:59models, but oftentimes what you'll see
5:01is they'll take Opus 55 and Fable 52 at
5:03similar reasoning efforts, give it the
5:05same tasks, they'll get the outputs
5:07back, the quality will be the same, but
5:09Opus will be usually half the price of
5:11Fable 52. So Fable 52 is 2x more
5:14expensive than Opus. And that's on the
5:15low end. Sometimes Fable 52 could be 3
5:17to 5 to 6x more expensive than Opus. And
5:20the reason being is that often times you
5:22don't want to compare just the price per
5:23token, which is what people often talk
5:25about. You want to you want to kind of
5:26compare price per task because an AI
5:29might spend a bunch of time in a bunch
5:30of tokens checking the output
5:32unnecessarily, which burns way more
5:34money than you would need for that given
5:36activity. And that's what Fable 52 might
5:38do if you set it to a higher effort than
5:40high, such as extra effort and/or max.
5:42And to that point, we've talked about
5:44model size and the associated cost, but
5:45effort is key here. So, usually when I
5:47work with people, again, they start with
5:49Opus 55 medium. And if the model hasn't
5:51achieved the task the way you wanted it
5:53to, and you have a good prompt and good
5:54context, then you can start a new chat
5:56and bump the effort level up to high.
5:58You don't change the model yet. You're
5:59just going to go from Opus 55i medium to
6:01high. In a new chat, you'll run the
6:02task. If it does it, great. If it
6:04doesn't, then that's when you're going
6:06to switch models. You're going to go
6:07from Opus 55i likely to fable 5.2
6:10medium. And the reason I recommend that
6:12is if you go to extra higher max inside
6:14of Opus 55 or even Fable, you're going
6:16to spend way more money than you need to
6:18for this given activity. And now what do
6:20I mean by more? Well, it's a lot more.
6:22Anthropic has done many tests themselves
6:24and they showed that when you use Opus
6:2655 and you go from high to extra high on
6:28the same task, you're likely going to
6:30spend around 2 and 1 halfx more on that
6:32activity. But the quality, which is very
6:34important, is identical or almost
6:36identical. So high is slightly lower
6:38than extra high when it comes to the uh
6:39quality of the output, but oftentimes
6:41it's not enough for you to actually pay
6:42that amount of money. So you're paying
6:44two and a halfx more, but your output is
6:46only minuscule and the quality that's
6:48improved. And one important thing to
6:49mention here on cost is if I take Opus
6:5155 at extra high reasoning, I give Fable
6:5352 at medium effort that same exact
6:55task, they're going to cost the same,
6:57but we're getting a huge jump in the
6:58quality and intelligence by Fable 52.
7:01And that's why I don't go above high for
7:02Opus 55 or Fable 5.2 do because we don't
7:05need to pay a bunch more for a quality
7:06that's going to be not that great. And I
7:08really want this to sink in, so I'm
7:09going to repeat this one more time. If
7:11you're using Opus or Fable, never go
7:13above high effort. High is as high as
7:15you should go. If you're going to go
7:16higher, you're going to burn a bunch of
7:17money unnecessarily for quality is just
7:19as good, if not worse. So now, when do
7:21we actually know when to use Fable?
7:22Well, for Fable 52, there are only two
7:24use cases that I found that are actually
7:26valid to use this tool for most
7:27knowledge workers. The first one is the
7:28most common. So in that case, we're
7:30actually going to have Opus 55 do the
7:31work and then Fable 52 will review the
7:34work. So let's say we take a series of
7:35source files and we give those source
7:37files to Opus 55 to do this task. After
7:39it's done the work, we then have the
7:40finished piece. We're going to start a
7:42new conversation and then we're going to
7:43give Fable 52 the finished work as well
7:45as the source files and ask it to review
7:47Opus 55's output. And I don't do this
7:49for every activity. I only do it for the
7:51activities that are high stakes. So
7:52those are going to be the activities
7:53that there's some sort of legal,
7:55reputational, or financial impact on the
7:57business. So this is the first situation
7:58where we're going to use fable. Again,
8:00this is the most common. The second
8:01situation is when we've actually given
8:03Opus 55 a chance and it failed.
8:05Actually, it failed twice. So in this
8:07situation, we're using Opus 55 at high
8:09and we're giving an activity and it
8:11failed. We then open up a new chat. We
8:12tried it again and it failed again in a
8:14similar way. So we can see that just
8:15this model doesn't have the intelligence
8:17to do this task the way we wanted to. So
8:18in that situation, we're going to start
8:20a net new conversation and we're then
8:22going to give Fable 52 probably at
8:24medium effort to do the task for us.
8:25Now, why do we start a new conversation?
8:27Two reasons. First is that you don't
8:28want to waste a bunch of money having
8:30Fable 52 reread the existing
8:31conversation of what Opus has already
8:33done. Because what people will sometimes
8:35do is they'll have Opus 55 do a task.
8:36They'll then change it from Opus 55 to
8:38Fable 52. When you do that in the same
8:40thread, it has to reread everything
8:42you've done. So, that's going to waste a
8:43bunch of money there, but also it's
8:45going to be biased towards what Opus 55
8:46has already done there. So, it's going
8:48to have some sort of bias towards that
8:49work and might give you a poor output.
8:51So instead, you want to start a new
8:52conversation, a fresh conversation with
8:54the source files and the exact prompt
8:55you want it to do. You then give it to
8:57Fable 52, let it do the activity for
8:59you. So we've talked about when to go up
9:01in model intelligence, but how do we
9:02know when to go down a model
9:03intelligence? The most common situation
9:05is when you have a recoccurring task. So
9:07let's say that you've built out a skill,
9:09the associated instructions and context
9:10with a higherend model. So we use Opus
9:1255, medium or high to do the buildout
9:14for us. Once we built out the skill, we
9:16set this as a scheduled task to do this
9:18process every single day for us every
9:19single week. and we've tested it
9:20multiple times with Opus 55 high then
9:23medium then low and it keeps doing the
9:25work perfectly. Then we can bump it down
9:27even further. We can go to sonnet 55
9:29high or sonnet 55 medium or lower and
9:32see if it can actually do the activity
9:33for us. We need to test it methodically
9:34on different inputs for this process
9:35that we're going to do time and time
9:36again. But if it proves itself and it
9:38gets it right every single time, we know
9:40that we can actually downgrade the model
9:41to save oursel money on this activity
9:43that's going to run every single hour,
9:44every single day, every single week. So
9:46now as a quick recap of everything I
9:48talked about here, the things you need
9:49to understand. First off, when you open
9:51up Claude, you should always start with
9:52Opus 55 medium on every single activity,
9:55especially if you're a knowledge worker.
9:56After that, if the activity fails for
9:58whatever reason and you have a good
9:59prompt and good context, then you can
10:01bump it up to high. But never go above
10:03high for effort because you're going to
10:04waste money on a quality that's just
10:05going to be the same as high. Once
10:06you've done that, if it's an activity
10:08that's critical and high stakes, you can
10:09then have Fable review the output from
10:11Opus. So again, you want to do that in a
10:13new chat. If opus has failed twice in a
10:15row in a similar way, then and only then
10:17do you actually use fable in a new chat
10:19on that activity. You're probably going
10:20to set it to medium effort first and
10:22then bump it up to high if necessary.
10:23And then finally, sometimes we want to
10:25downgrade the intelligence to a lower
10:26model so we can save on money. And in
10:28that situation, you've likely already
10:29built out the instructions, the skill,
10:31the scheduled tasks with a smarter
10:32model. And then after you built it out
10:34and tested with a smaller one, you then
10:35can run it on a continuous basis every
10:37hour, day, or week. And that's it. So as
10:40a reminder, two quick things. First off,
10:42I have a 30-day AI insight series. It's
10:45completely free. You'll get 30 insights
10:47in your inbox of how you can apply AI to
10:49your business and your work. And the
10:50second thing is if you're paying for AI
10:53and you want your team actually using
10:55it, there's a link below to grab a free
10:57call with me as well as the offering
10:58associated to that. And if you've
11:00enjoyed this video, the YouTube gods
11:02know that you're going to love this
11:03video, so you should give it a click.
11:05Let's see if they're right. And I'll see
11:06you next time, internet.

About this transcript

This page contains the full transcript of Claude Fable 5.2: Anthropic Says Use Opus for Most Work by Dylan Davis, generated from the public captions YouTube serves with the video. The transcript has 2,807 words across 375 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.