YT YouTube2Text

YouTube2Text → Transcripts → Anthropic lanza Opus 5.5: lo probé contra GPT-6 Astra

Anthropic lanza Opus 5.5: lo probé contra GPT-6 Astra — Transcript

by Benjamín Cordero · 4,815 words · 690 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Anthropic just launched Opus 5. GPT6
0:05Astra. In their table, we can see it
0:08beats it in several programming tests
0:10and professional work. And the cost
0:12shown here, I think, is a huge part of
0:15this entire new announcement. This is
0:19an excellent model, but Astra is still
0:21ahead in two benchmarks that still
0:23matter to us, which are Business
0:24Workflow, right here, and Agentic
0:26Scientific Research. But now we can see
0:30that Opus is crushing it in everything
0:32related to coding, and it's doing very
0:35well in many other tests as well, even
0:38omitting comparisons with Astra in some
0:40cases. So, the question is, should I
0:43switch? Has Anthropic really returned
0:45to being the king? We are going to open
0:49the announcement, break it down from
0:51top to bottom, and analyze all the
0:52prices, the benchmarks, what each one
0:54means, and my personal take—whether I
0:56ended up liking it or not. We are going
0:59to create a page with images from GPT
1:01Image 2.5. We will connect it to
1:04Highfield and Sidans 2.5, and we will
1:06ask for different tasks to see if it
1:09truly has the same level of performance
1:11between GPT6 Astra, for example, and
1:13Opus 5. Before we start, I would really
1:17appreciate it if you could leave a like
1:19on this video, not only because it
1:21helps me, but because you also tell
1:23your algorithm that you like this style
1:24of content and it starts recommending
1:26more and more of it. Now, let's get
1:29down to business. Okay, before we begin
1:32, let's set two prompts running so that
1:34at the end of this video we can see the
1:36result and compare GPT6 Astra with Opus
1:385.5. I’m going to throw the same
1:42prompt at both of them, which is to
1:43build a page—a landing page—
1:44connected to Highfield using Sidan
1:46Caches 2.5. And well, we'll see what
1:49happens, and then we'll analyze which
1:51one was actually the best. So, here we
1:53can already see Opus 5. Max. And now
1:57I’m going to go into Codex and ask
1:59for the exact same thing, both at their
2:02highest reasoning levels. Okay, now
2:06let's see that, uh, a few seconds or
2:09minutes ago, Anthropic made this launch
2:12, which was their new model, Opus 5. It
2:18says it performs at the level of Claude
2:21Fable 5.1. One, but this is the
2:25interesting part: it costs 40%less to
2:28run than Opus 5, which is truly a
2:30stroke of genius considering new models
2:33like, uh, JEV, which are running very,
2:36very cheaply. And now we are starting
2:40to compete; while they aren't the same,
2:42we are definitely starting to compete
2:44quite a bit on price, because
2:45alternatives that are much, much
2:46cheaper are starting to appear. But why
2:50don't we see what this really means?
2:52And what is, what is the whole point of
2:55this launch, because I started reading
2:57and it looked quite interesting. Here
2:59we have the main benchmarks comparing
3:02them to Opus 5. With the previous
3:06version, Opus 5, Claude 5.1, and GPT-4o
3:10. And well, what I also want to
3:13highlight is that notice here they say
3:16it is the first model in their new 3.5
3:18family. In other words, Sonnet and
3:22Haiku could probably be coming in the
3:23next few weeks, or perhaps they'll
3:24discontinue Haiku. I don't know, let's
3:27see what happens, but there will
3:29undoubtedly be a new Sonnet or maybe
3:31even a Claude. Uh, from what I read, it
3:35functions quite similarly in reasoning
3:38levels to Claude 3.5, and this is their
3:41official launch blog. Uh, what I want
3:44to see here, well, we already know what
3:47we always do, let's translate it to
3:49English so there are no differences or
3:52problems with language barriers. And
3:55let's see the funny translations that
3:57Google Translate gives us here. Uh, and
4:00notice they tell you that in most tasks
4:03, its operation costs 40%less than Opus
4:063. And notice this comparison they are
4:09showing us here, uh, because further
4:11down we will see how much it costs
4:13according to Astra and GPT-4o. Uh, they
4:17announce a different generation of
4:19responses. They tell us that responses
4:22are 30%faster than Opus 3, uh, and they
4:24tell you that in addition to the price
4:27drop, they increased the usage limit to
4:295 hours on Pro, Team, and Enterprise
4:31plans. So, they raised the limits to
4:34these 5 hours. Great. Uh, and notice
4:37that they also told us here that they
4:40gave us a, uh, token reset, where I
4:42found this here in their post. Uh,
4:47besides increasing the 5-hour usage
4:49limit, they also gave people, or users,
4:52or subscriptions, a rate limit reset
4:54that can be used whenever you want.
4:59This is something Codex already had,
5:01which I found great because, basically,
5:03you have periods of higher intensity or
5:05work where you need these resets, so I
5:07think it was an excellent, excellent
5:09success. We never really know how much
5:13a subscription yields for us, so they
5:14can tell us yes, eh, it has a higher
5:16limit, but in the end we can never
5:18really verify it because this is
5:19something they adjust based on
5:21computing demand and the supply they
5:23have. If we start looking here, well,
5:26it communicates more naturally, blah
5:28blah blah blah. Right, yes, obviously,
5:30I’m always going to want to say what
5:32suits me best here. So let's go down to
5:36this, let’s get into this, which is
5:39what really interests us, which would
5:41be the different benchmarks between
5:43Fable 5.1 and the rest of the
5:45benchmarks. Eh, we’re going to do
5:49what we always do, take a screenshot so
5:51we can write on or show things. And
5:54let's see what each one means. Here we
5:57have a series of benchmarks that are
6:00standardizations carried out by
6:02different people or tests used to
6:04measure all these models in an attempt
6:07to have some kind of criteria for doing
6:10so. And Anthropic itself, in fact,
6:15tells us here that these benchmarks are
6:17even becoming less and less
6:19representative because something also
6:21happens here where they isolate the
6:23models from their harnesses. So, it's
6:27not exactly the flow one has when
6:29working in something like Claude Code
6:32or Codex, eh, because the model is
6:34being isolated a bit, but anyway, it is
6:36still an excellent, excellent parameter
6:39. The first one that caught my
6:43attention is this one here, agent
6:45coding or agentic coding, which is the
6:48Terminal Bench 4.0, which tells us in
6:51the end how well an agent can complete
6:53tasks from a terminal. I'm sorry, I
6:57have to take a screenshot because I
6:59really need to see this in English,
7:01otherwise it’s very hard for me, eh,
7:05to be able to read the numbers and
7:07things well here. So now, yes. Eh,
7:10agentic Terminal Bench 4.0. We have
7:13Opus here which scored 66%versus its
7:15previous version which was 52%. So, it
7:20won here and it also beat Astra by
7:22quite a bit, which is super good, eh,
7:24because Astra was already super good at
7:27this. I personally used it quite a bit,
7:31but here Opus 5. So, let's think of
7:39this benchmark as the ability to solve
7:41a task through the terminal, that is,
7:43where the computer’s actions are
7:46executed. It does it better than all
7:49the other models and would be the new
7:50state of the art. Next, we have another
7:54benchmark here that is also about
7:56coding, but I believe in the one above,
7:58Terminal Bench 4.0, they only ask a
8:00single coding-related question that is
8:02also solved. This one involves a series
8:06of more steps; I mean, it's like a
8:08problem, a project, it involves more
8:10steps unlike Terminal Bench 4. Frontier
8:13Code goes beyond just answering a
8:16single question, and here it gets 54%
8:18versus eh Fable which gets 53%, meaning
8:21it beats it and shows a substantial
8:24improvement over Fable 5.1. I can't
8:27wait for Fable 5.5 to come out because
8:29it's going to be very interesting, as
8:31it will surely show us significant
8:33improvements as well, without a doubt.
8:36Eh, next we have another one here for
8:39Agentic Coding again, and this one is
8:41Cursor Bench 4.0, which also performs
8:43better than Fable 5.1. But it intrigues
8:47me because they aren't measuring Astra
8:49here, but in the end, this is a task,
8:51it's also Agentic Coding, but it's done
8:53in the Cursor environment; that's why
8:56it's called Cursor Bench 4.0. Here it
9:00performed with almost 58%. And let's
9:04remember all this is based in the end
9:06on the sessions you have within the
9:08Cursor harness, which allows you to use
9:11different models, which is the
9:12interesting part. Yes, it intrigues me
9:16quite a bit why the Astra box isn't
9:18there. I'm going to remove this here.
9:20And this one too. Eh, it intrigues me
9:22quite a bit why it's not there, because
9:24obviously, it not being there doesn't
9:25mean zero. Next, we go to this other
9:29test which would be knowledge work, eh,
9:32which is quite the opposite of a
9:35percentage, sorry, but it is measured
9:39in points. And here we have 100 versus
9:42the others, eh, where Fable 5.1 was
9:45already at a super high number and GPT-
9:474o Astra wasn't as high. This is a
9:52benchmark that evaluates professional
9:54work from I don't remember the exact
9:57number of professions, but basically,
10:00it asks how good the outputs of this or
10:03that deliverable are in the end, eh?
10:08And then it assigns a score, and that
10:10score for Opus 5...Is 1846, and for the
10:12other models, it is lower. Here I was
10:17also a bit surprised that GPT-4o Astra
10:20reached 1542. Eh, again, this is a kind
10:25of score like the one used in chess or
10:27an Elo rating. Eh, I don't know, higher
10:31is obviously better. This is a
10:34benchmark that I am quite interested in
10:36, and I think it is one of the most
10:38important ones for people who build
10:40automations: the Business Workflow
10:42benchmark. Uh, this is a benchmark that
10:45Zapier put together, and it ultimately
10:47measures how well it connects with the
10:49different applications we use in our
10:51day-to-day business. I mean, it takes a
10:55task and then measures whether, uh,
10:57people can correctly complete that
11:00specific task. This is much closer to
11:03what is actually done and how we use it
11:04in our day-to-day lives. I mean,
11:05especially for people who run
11:07businesses. It's like looking for
11:09something in a CRM or, I don't know,
11:11updating a record in a database or
11:13contacting and following up with this
11:16person here. uh, coordinating actions
11:18between different tools with different
11:20connections, and so on. Yeah, I think
11:22this is an excellent benchmark. And
11:24here it gets 40%and it is beaten by GPT
11:27-6 Sol, uh, sorry, GPT-6 Astra with 41%
11:30. Side note: it wouldn't surprise me if
11:33GPT-6 Sol comes out while I'm uploading
11:35this video. I think it's very likely
11:37they'll do something like that, so we
11:39just have to wait. But, uh, here GPT-6
11:43Astra does win, but where Opus 5 really
11:46wins, uh, by quite a lot, rather. Is in
11:50the Humanities Last Exam. And this is
11:54an exam that I also like quite a bit
11:57because it ultimately gathers many
11:59questions or answers from different
12:02disciplines, sort of like a PhD, so to
12:05speak, and takes these specialists and
12:08then gives them a percentage, let's say
12:11, of, uh, evaluation in specialized
12:14knowledge. Uh, and here they measure it
12:17with tools, which is a bit more
12:19real-world. In fact, without tools, I
12:21don't really understand why it's done
12:23so much, since we are using tools in
12:25our day-to-day life anyway, but it's
12:27okay to isolate it; it's interesting.
12:30Uh, but yes, Opus 5 wins here. GPT-6,
12:35uh, Astra gets 67.7%versus 57.2%. So,
12:43here both results are with the tools.
12:46Obviously, uh, getting a sort of good
12:49grade in the end on an exam doesn't
12:52mean, uh, that it's the same as doing
12:55autonomous research in a certain way,
12:58but anyway, I think Opus 5 performs
13:01much better here. Than GPT-6 Astra. Uh,
13:07and the other category where GPT-6
13:09Astra does beat Opus 5. Is the
13:14Scientific Research benchmark, which is
13:16also a terminal-based science benchmark
13:17. And here, scientific research tasks
13:20are evaluated based on an agent working
13:22in a computer terminal. This Astra did
13:26come out on top with 64%versus Opus 5.
13:30Which was at 58, almost 59%. But anyway
13:35, the announcement does report standard
13:38deviation errors at several points at
13:41the end, so this could obviously still
13:44change, but it gives you a slight idea
13:47of what we're actually facing. This was
13:51the one I wanted to get to, which is
13:53OSWorld 2.0. Why? Because GPT-4o stood
13:57out quite a bit historically for having
13:59excellent computer use. Uh, and here
14:03Opus 5 is at 1.8%. Meaning, it beat or
14:08performed better than its previous
14:12models, but it isn't comparing it to
14:15GPT-4o Astra. We'll take a look at this
14:19test, but uh, it measures how well an
14:22AI navigates an interface designed for
14:25humans, which is where GPT-4o Astra was
14:28performing super well. If I go here and
14:33go to GPT-4o Astra and look for this,
14:37like their launch page, I'll be able to
14:42find here that they get a, uh, 72.6%.
14:47But look, they compare Opus here with
14:4970, Opus 5, obviously, with 70.2%. But
14:54if I come back here, it says Opus 5 has
14:5774%, so, like, a little bit more. But,
15:01anyway, here we are comparing, if we
15:04had to compare computer use, that is,
15:06how well an AI performs navigating
15:09interfaces designed for humans. Uh, GPT
15:13-4o Astra has 72.6%, Opus has 70, Opus
15:185 and the new model, uh, Opus 5. It is
15:23quite a bit higher than that, I mean,
15:26it is six or seven points higher than
15:29Opus 5. So, seven points higher than
15:33Opus 5 here would be, well, at 77%
15:37versus 72%. So it would still be
15:40winning. Obviously, we aren't using the
15:43same parameters here, the reference
15:45isn't exact, but it tells us they
15:47focused quite a bit on computer use,
15:49which is what made many of us start
15:51using GPT-4o Astra. So it's good to
15:54know that Opus has taken the lead again
15:55, at least here in everything related
15:57to computer use. And then we have the
16:01recognition of visual graphics, which
16:04is this series of, like, cartography,
16:07which also isn't compared to GPT-4o
16:10Astra, but we can see this is basically
16:13an interpretation of how well the
16:16models understand plans, axes,
16:19distances, a bit more thought out
16:21especially since it started being
16:24implemented with Blender, Rhino,
16:27AutoCAD, and all. all these things. Uh,
16:30it’s like a mapping test, so to speak
16:32, an understanding of space and the
16:34physical plane where it performed at 89
16:37%and 84%, in other words, 89%versus
16:4288.4%and 83%before. Again, we don't
16:45have the reference for GPT6, but well,
16:47let’s see how much it costs us to get
16:50it. Here they tell us, "The main
16:52advantage of Opus lies in its
16:54efficiency; it costs less per token
16:56than Opus 5 and uses fewer tokens per
16:58task, which is a 40%reduction in costs.
17:02Uh, well, a token is this small unit of
17:05measurement used to standardize the
17:08inputs and outputs that an AI has.
17:12It’s basically how this text is
17:14separated and it would be like gasoline
17:16; spending more tokens will only
17:18increase your bill. Here is Opus 5. It
17:24costs $ 4 for input and $ 5, uh, versus
17:28the $ 5 of Opus 5, and $ 20 for output
17:32versus $ 25 for the output of Opus 5.
17:37So, here for input it went down and for
17:39output it went down $ 5. Now, here, of
17:44course, it makes you a bit curious
17:47where this 40%comes from, if this isn't
17:5040%, this is only a one-fifth decrease,
17:54right? When you combine lower rates
17:58with lower consumption, in the end, you
18:00should arrive at this kind of 40%.
18:04Because it’s not only cheaper, but
18:07they also improved the cache writes
18:10here. So, in the end, all of this
18:12compounded is what makes us reach that
18:14figure. Look, uh, they also tell you
18:17that fast mode is available in Cloud
18:20Code and the platform with a speed up
18:22to 2.5 times, uh, and it costs you
18:24double, basically, just like all the
18:27others. The interesting thing, though,
18:31is if we compare it later with GPT6
18:33Astra and we look here at what would be
18:36the price, uh, which is $ 10 per
18:38million input tokens and $ 50 per
18:40million output tokens, versus these
18:43which would be 4:20, it would be up to
18:4560%less or more economical than Astra
18:48for better performance. So, it clearly
18:52depends on your field. Obviously, if
18:56you are involved in biology, maybe, or
18:58scientific research, you will want to
19:00keep using Astra, but for most tasks,
19:02uh, Opus 5.5 should work better for you
19:04. Uh, next here we have coding, we can
19:09see that it focuses on migration and
19:11audits of large projects. For example,
19:15here 200,000 lines in 3 hours. Before,
19:18it took more than 20 hours and consumed
19:192.5 times more tokens in an internal
19:21run. They did the same thing and
19:24managed to do it, well, we already saw,
19:26in less than 3 hours, I mean, from 20
19:28hours to less than 3 hours. Anyway,
19:30this is a case documented by them, uh,
19:32anyway, but look at this. Here we have
19:35what we already saw. Look at the curves
19:37, how they show and compare them to us.
19:40Grey being GPT-4o Astra and Opus 3.5
19:42being, obviously, the orange one here.
19:46On these curves, upwards you have a
19:47higher score, towards the right you
19:49have more cost, towards the left you
19:50have less cost per attempt. Obviously,
19:55this is also represented in a
19:57logarithmic way, I mean, it's not that
20:00it's proportionally more money, at
20:02least on the cost axis, but in the end,
20:05it really helps us visually represent a
20:08difference in the changes, at least.
20:12But what they want to show us here is
20:15that Opus 3.5 equals GPT-4o Astra in
20:18the terminal bench, in average
20:20reasoning, when it's at its maximum, so
20:23to speak. And if we look at the cost,
20:28the best attempt by GPT-4o Astra to get
20:31a 58%spent $ 7, and the attempt for
20:35almost 58%spent $ 3. I mean, here we
20:39have a reduction of practically half. I
20:43mean, we can do terminal bench tasks,
20:45all the agentic coding stuff, at half
20:48the price, achieving the best result of
20:50GPT-4o Astra, which is super good, at
20:53least to keep in mind if you are
20:55building things with AI, which is what
20:57matters most to us, right? Then, well,
21:01all of this is coding; here, the
21:03Frontier code also looks much better,
21:05especially in terms of price. You know,
21:08look at this. Here we have 40 cents for
21:11Opus on low, and here we have uh 1.5.
21:18Dollars, I mean, three times less for
21:21GPT-4o Astra while getting a lower
21:23result, and obviously, if we get to
21:26maximum reasoning, we get to similar
21:29things in this sense. But my point is
21:32that you can see the trend a little
21:35closer here, much more economical. The
21:37best point at 54%and the other at 53%,
21:41but spending $ 4 versus 80 cents, I
21:44mean, five times more. That's why it's
21:48also worth looking at this curve,
21:50because it helps us understand the
21:52relationship between costs and uh very
21:54well. the scores or the outputs. The
22:01following sections, in the end, show us
22:03how they improved in terms of security,
22:06uh, in everything related to prompt
22:09injection, which, well, is a super
22:11important part if you don't want to end
22:14up using your McDonald's chatbot as an
22:16LLM, uh, and obviously if you want to
22:19keep things secure. Then, if we keep
22:23scrolling down, we can find this test
22:25here that we also saw, moving a bit
22:27away from code. We're going to start
22:30entering reports, analysis, and
22:31presentations. It has an overall
22:33performance or score, or the Elo, which
22:36would be this score we saw earlier,
22:39significantly higher than all other
22:41models. Uh, if we keep scrolling down
22:45here, they start showing us different
22:48cases, uh, different testimonials on
22:50how it changed versus its previous
22:53models. For example, here both models
22:57are comparing an error or explaining a
23:00problem with an error that occurred in
23:02billing. Uh, so here we can see that it
23:05starts with the consequence, helping
23:07you understand what the error is.
23:11Unlike here, it explains the error
23:13first and then it sort of starts
23:14explaining the consequence, from what I
23:16understood from what I managed to read
23:19here. If we keep scrolling down, we
23:22will find more things regarding
23:23security. Yes, we keep scrolling down
23:26here. Obviously, more security stuff
23:28regarding alignment and, well, all
23:30these things that you can start looking
23:32at if you're also interested in how it
23:34works and how all the APIs connect. Uh,
23:38here, well, they restricted or
23:40unrestricted, so to speak, the capacity
23:42or the blocks that had been placed on
23:45biology. Uh, it even outperforms
23:48Mixtral 5.1 in many areas of work. And
23:51over here we already have the
23:53availability. It is available on all
23:55platforms, so to speak. If you go into
23:59Claude, you will find this here. I
24:02loved that they give you a limit to
24:04reset it. On October 22nd, it will also
24:06be there until October 22nd. This means
24:08you have a sort of reset, the same as
24:10what Codex had, uh, that you can use
24:11during your most intensive periods. And
24:13now, let's see how it performed here,
24:15since it finished. Uh, let's see the
24:18final results. Here they spent, or
24:21rather, it took...Let's see. How long
24:23did it take? That's good. Uh, we sent
24:27the prompt 40 minutes ago and it
24:29finished 3 minutes ago, so it took, uh,
24:32something like 35 minutes. This would
24:35be Opus. And over here, uh, it
24:37delivered 19 minutes ago, and we sent
24:39it at the same time. It took 21 minutes
24:42, I mean, Opus 5. It took longer than,
24:46uh, GPT6 Astra. But let's see the
24:49result, which is what matters. Here we
24:51are going to open it with, or actually,
24:54I’ll just tell it to open it in
24:56Chrome to see it. I’ll hit enter, and
24:59here I’ll also tell it to open it in
25:01Chrome to see it. Uh, what we asked it
25:04to do was create this page of some
25:06exploded-view headphones. Uh, we were
25:09also connected to Highfield. Highfield
25:12is the processor we use to be able to,
25:15uh, centralize all the AI models in one
25:19place. It aligns very well with
25:22everything we do because, in the end,
25:24it allows us to have everything,
25:26obviously, in one place, including the
25:28latest AI models. As we can see here,
25:30if I go to images, I can create images
25:33with several models just by uploading
25:35images to the same place. So, of course
25:38, it came in here, generated the images
25:40, then the videos, right? It generated
25:42the exploded view with Sidans 2.5,
25:44which were the things. Uh, this
25:47Genjutsu one is also super interesting;
25:49it’s for changing your position, or
25:51changing your shirt or some outfit or
25:54whatever. And obviously, it has MSP. We
25:58connected via MSP to Cloud Code and
26:00asked it to create the pages for us,
26:02and this is what it gave us. This right
26:06here is the page from, uh, GPT6 Astra.
26:10As we can see, uh, it works really well
26:12. The exploded view is quite good. At
26:15the end, there’s just a small bug
26:17where it kind of transitions to this
26:19new thing, let's say, or it kind of
26:21explodes, or rather, it becomes a bit
26:24transparent. But overall, I think
26:26it’s super, super good." Sobora audio
26:28, your world on pause, "it shows us.
26:30There’s a small bug here too, uh, but
26:33the rest works well in the end; it gave
26:35us the exploded view, it gave us the
26:37detail, and the page is responsive.
26:40Obviously, we don’t want it to be, uh
26:43, public, right? Because, in the end,
26:46it’s not something we’re interested
26:47in. This has been a pretty good detail,
26:49as it maintained consistency really
26:51well for us. This is the page that GPT6
26:53Astra built for us. Let’s see the one
26:55Opus 5 built for us. Wow, it looks
26:58pretty good, huh? Let’s see if, well,
27:01it’s obviously not controlling the
27:03browser here. I don’t want it to
27:05control it for us, so we’ll just open
27:07it here because I don’t want it
27:09controlling my browser. Uh, and it
27:11looks pretty good and, wow, wow, if we
27:13start scrolling down. Okay, okay, if we
27:16start going down, I think, uh it did a
27:20very good job, at least at first glance
27:24. Like, look, it's like a little
27:26presentation that moves forward, like
27:28with the view, wow, wow, it's super
27:31good. I think, look, it even gave us
27:33this image here with icons. This video,
27:36uh, reserved in Obsidian, it gave us
27:38the different colors, and well, I think
27:41we have a clear winner here, at least
27:43in feeling. I have to say that this is
27:46the best page I've seen with such a
27:48simple prompt. I'm not exaggerating, I
27:50think it's the best. Uh, wow. Let's see
27:54the responsiveness if we close it here.
27:57Wow, this is really impressive. It's
28:02the best page I've seen with such a
28:04simple prompt, you know? I mean, we
28:07didn't even manage to let it finish. In
28:10fact, I interrupted it ahead of time
28:12because it kept doing, like, the whole,
28:14the analysis, but wow, wow. I have to
28:17say it's actually quite good, honestly.
28:19Uh, congratulations to Claude. I think
28:21it did a better job here. It took
28:22longer. Yes, twice as long, but
28:24basically I don't care if it takes
28:26longer, I care about a better result.
28:29If I wanted to use cheap models, I'd
28:31use others like Jeff, which by the way,
28:33I recommend you stay tuned because
28:34something very interesting is coming
28:36with how I'm using it today. This is
28:39excellent. It's excellent. Well, anyway
28:42, uh, returning to the initial question
28:44, is it worth it for us to switch from
28:46GPT-4o to Opus? In my opinion, if
28:50you're not into scientific research or
28:53biology, chemistry research, or very
28:55specific things you can see there, the
28:58answer is yes. I think the Opus 3.5
29:02announcement in the end was super good,
29:05super relevant. Uh, it obviously
29:08deserves a serious test if you do
29:09things like long programming, or tasks,
29:11or reports, or whatever, anything
29:13related to design. Look at this. Uh,
29:17and obviously complemented with
29:18Artifacts, it works super well because
29:20we can get these kinds of things all in
29:22one place. So it's epic, it's epic. Uh,
29:28regarding business automation, I think
29:30it works quite well. I still have to
29:32test it a bit more. If you want me to
29:34make a video below comparing GPT-4o
29:36with Opus 3. Like in various use cases,
29:39let me know. Uh, really, I, uh, read
29:41all the comments, I respond to all the
29:44comments, uh, I really value it when
29:46you leave me comments. And it's not
29:48just me; I also have a scraper that
29:49pulls comments from my videos and
29:51provides feedback, like:" Oh yeah, look
29:53, a lot of people are asking for this
29:54video, "and then I create it. So, it's
29:57also really helpful when you leave
29:59those things, you know? So well, Opus
30:02is back, which makes me very happy
30:04because I've always been a fan of
30:06Claude Code. I still need to test it a
30:09bit more, perhaps on the computer use
30:10side of things. If you're deciding
30:12between Codex or Claude Code, on which
30:14one to learn. I have courses for both,
30:17completely free. You can find them on
30:19YouTube, and if you want to find out as
30:21soon as all these things come out, if
30:24you look here at Imperio Argéntico, we
30:26posted it as soon as it dropped—I
30:28mean, I posted it an hour ago. It
30:30already has 140 comments. We are all
30:32super happy there. Obviously, I'm going
30:35to upload this video, so yeah, many,
30:37many good things are happening inside
30:39Imperio, and this is where we find out
30:41about everything long before everyone
30:43else so we have a competitive edge.
30:46Plus, we obviously have live sessions,
30:48we have the courses, we have everything
30:50. It's really good, it's really good.
30:52Stop by if you're interested in AI and
30:54are serious about building with AI.
30:57That said, I hope you liked this video.
30:59See you soon.

About this transcript

This page contains the full transcript of Anthropic lanza Opus 5.5: lo probé contra GPT-6 Astra by Benjamín Cordero, generated from the public captions YouTube serves with the video. The transcript has 4,815 words across 690 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 translate the transcript into 100+ languages.

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, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.