Экстракт Знаний
HUGE Claude Code Update: Opus 4.5, Desktop App + More!
11:24

HUGE Claude Code Update: Opus 4.5, Desktop App + More!

Ray Amjad 25.11.2025 11 016 просмотров 225 лайков обн. 18.02.2026
Смотреть на YouTube
Поделиться Telegram VK Бот
Транскрипт Скачать .md
Анализ с AI
Описание видео
Level up with my Claude Code Masterclass 👉 https://www.masterclaudecode.com/ Learn the AI I'm learning with my newsletter 👉 https://newsletter.rayamjad.com/ Got any questions? DM me on Instagram 👉 https://www.instagram.com/theramjad/ 🎙️ Sign up to the HyperWhisper Windows Waitlist 👉 https://forms.gle/yCuqmEUrfKKnd6sN7 Since I've never accepted a sponsor, my videos are made possible by... —— MY CLASSES —— 🚀 Claude Code Masterclass: https://www.masterclaudecode.com/?utm_source=youtube&utm_campaign=_kaSD-W56wE - Use coupon code YEAR2026 for 35% off —— MY APPS —— 🎙️ HyperWhisper, write 5x faster with your voice: https://www.hyperwhisper.com/?utm_source=youtube&utm_campaign=_kaSD-W56wE - Use coupon code YEAR2026 for 35% off 📲 Tensor AI: Never Miss the AI News - on iOS: https://apps.apple.com/us/app/ai-news-tensor-ai/id6746403746 - on Android: https://play.google.com/store/apps/details?id=app.tensorai.tensorai - 100% FREE 📹 VidTempla, Manage YouTube Descriptions at Scale: http://vidtempla.com/?utm_source=youtube&utm_campaign=_kaSD-W56wE 💬 AgentStack, AI agents for customer support and sales: https://www.agentstack.build/?utm_source=youtube&utm_campaign=_kaSD-W56wE - Request private beta by emailing r@rayamjad.com ————— CONNECT WITH ME 🐦 X: https://x.com/@theramjad 👥 LinkedIn: https://www.linkedin.com/in/rayamjad/ 📸 Instagram: https://www.instagram.com/theramjad/ 🌍 My website/blog: https://www.rayamjad.com/ ————— Links: - https://www.anthropic.com/news/claude-opus-4-5 - https://x.com/claudeai/status/1993030546243699119/photo/1 - https://x.com/ImSh4yy/status/1993030994480578952/photo/1 - https://x.com/alexalbert__/status/1993030687881080944/photo/1 - https://assets.anthropic.com/m/64823ba7485345a7/Claude-Opus-4-5-System-Card.pdf - https://scale.com/leaderboard/swe_bench_pro_public - https://x.com/danshipper/status/1993034359088050394 - https://x.com/jekbradbury/status/1993036318830739709?s=12 - https://x.com/pallavmac/status/1993059671780057403 - https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md Timestamps: 00:00 - Intro 00:26 - Opus 4.5 Introduction 01:08 - Trying Opus 4.5 on a Hard Problem 03:09 - Claude Desktop App 04:30 - Opus 4.5's Solution 06:14 - Opus 4.5's Output Tokens 06:38 - Opus 4.5 on Ultrathink 07:37 - What X is Saying 08:40 - Opus 4.5's Design 09:01 - Extra Usage 09:41 - Readline-style ctrl+y 10:00 - Starting Background Tasks with & 10:46 - Conclusion

Оглавление (13 сегментов)

  1. 0:00 Intro 123 сл.
  2. 0:26 Opus 4.5 Introduction 147 сл.
  3. 1:08 Trying Opus 4.5 on a Hard Problem 454 сл.
  4. 3:09 Claude Desktop App 318 сл.
  5. 4:30 Opus 4.5's Solution 363 сл.
  6. 6:14 Opus 4.5's Output Tokens 93 сл.
  7. 6:38 Opus 4.5 on Ultrathink 210 сл.
  8. 7:37 What X is Saying 256 сл.
  9. 8:40 Opus 4.5's Design 73 сл.
  10. 9:01 Extra Usage 144 сл.
  11. 9:41 Readline-style ctrl+y 81 сл.
  12. 10:00 Starting Background Tasks with & 162 сл.
  13. 10:46 Conclusion 178 сл.
0:00

Intro

Okay, so we have perhaps one of the  biggest Claude Code updates yet,   which includes a brand new  model, Claude Code for desktop,   and a bunch of other things that we'll be going  over in this video. But before getting started,   it's worth mentioning that I have a Black  Friday sale going on my Claude Code masterclass,   which will be linked down below with a  coupon code as well. It's probably the most   comprehensive resource you'll find on Claude  Code outside of the official Anthropic docs. And in the masterclass, I do share a bunch of  bonus content and strategies that I haven't shared   before on my YouTube channel. Anyways, first of  all, we now have Opus 4. 5, which is three times
0:26

Opus 4.5 Introduction

cheaper than Opus 4. 1 was at $5 per million input  tokens and $25 per million output tokens. And I   think that finally makes it usable for some people  and means that you won't hit your rate limits   quite as much. It is now the best coding model  in the entire world on SWE-bench Verified and   also on SWE-bench Pro, which is an even harder  version of this benchmark made by Scale AI. So they haven't added it to a list on the  right hand side here, but you can tell it   will be the best model in the world once it  is added. And one thing that stood out to me   is that now it's novel problem solving ability  via ARC-AGI-2. I think is now best in the world. So beating Gemini 3 Pro and also GPT-5.   So I actually want to try this out myself
1:08

Trying Opus 4.5 on a Hard Problem

to see how well this performs. So  there's been a pretty persistent   bug with my application HyperWhisper that I  haven't been able to solve for some users. There's a Black Friday sale going on  right now for the application as well,   linked down below. And some people have  switched over to application from bigger   competitors because they just prefer it. Anyways,  the persistent bug has been like this one over   here affecting 51 users, and I have no idea  what has been causing this bug for people. So what I did is I got GPT 5. 1 max high from  Codex CLI, that's a mouthful, to basically look   for Sentry and figure out what the problem is  here. And it proposes a few different problems. So after about three or four minutes of  thinking, it came up with this. This was   pretty useless. Then I tried again, it  came up with this, also pretty useless. Then I tried again with this, also  pretty useless. And I was like, okay,   well, clearly a GPT 5. 1 Codex,  this is a pretty novel problem. It's not solving it. So then what I did  is I gave the same problem and the events   from Sentry and the HyperWhisper code to  to GPT 5. 1 Pro and after 41 minutes of   thinking it basically came up with a few  set of possibilities and I was honestly   pretty surprised by the result over here  because it identified problems that GPT 5. 1 Max High was not able to identify and then gave a  bunch of solutions for them as well. And you can   see like all its thinking over here. And this is  probably the longest time that I've gotten GPT 5. 1 Pro to think for so far. So I want to  basically see if Opus 4. 1 is able to   come up with a pretty similar solution or if  it identifies something completely different. So basically I'm actually going to download  the same Sentry events that I gave it. And   you want to make sure you're on  the latest version of a Claude,   so just run a claude update. And I  think right now it's version 2. 0. 52. And then switch to Opus 4. 5, which is now the  default recommended model. So press enter here   and then give it the same file and then copy over  the same prompt that I gave the other models,   press enter, and then we'll see if  it actually finds what the issue is. And I suspect that it may be able to do so  because a novel problem solving ability is   like pretty good here. Anyways, whilst we're  waiting to see if it gets it, there is now an
3:09

Claude Desktop App

update in the Claude desktop app. So if you don't  already have that installed, we now have Opus 4. 5 in the app, and we also now have Claude  Code, the like web version of Claude Code   built in. So if you go to sidebar here, and  then previously, if you just clicked on code,   it would redirect you to the web version of the  application. But now if you update it, now you   should see a toggle over here, and that allows  you to run Claude Code in a preview version. So install runtime dependencies. Then  you can either run a session on your   computer with a local git worktree.   And if you don't know already,   git worktrees basically allow you to have  multiple git branches at the same time. So you can work on multiple things independently   without having to duplicate the folder.   Or you can have it running on the cloud   with any of these environments that you create  here. And then you can resume that with your   computer or your phone from the Claude  web application or the mobile application. Anyways, let's kick off a task here as well in a  brand new worktree. And then using HyperWhisper,   I'm going to say, can you find a bug in  this code base and then go ahead and fix   it? And that will transcribe what I just  said and then press enter and that will   kick off in a local worktree. And then  basically if I show the hidden folders,   I can see a worktrees folder and then I  can see what it's coding inside of here. Anyways, this is pretty nice and  I will be using it more often,   I think. I didn't like having big sessions running  on the cloud version of Claude Code because I like   having like big tasks running locally instead.   Anyways, going back to what we gave Opus 4. 5
4:30

Opus 4.5's Solution

earlier, the same problem. You  can see it's used about 44,000   tokens so far and that's probably  with my CLAUDE. md file as well. And then basically it identified these problems  over here and it seems to have identified some   of the problems that GPT 5. 1 Codex Max High was  not able to identify but let me just do a quick   comparison of this to GPT 5. 1 Pro and what I was  able to identify because I think the main thing   the app came modal loop is actually the same thing  okay so I basically got Nano Banana Pro 4k to   generate a comparison between the two solutions  and solution one is the one that opus 4. 5 came up with and solution two is  GPT 5. 1 Pro via the website but   yeah I don't think this diagram is actually  that useful so let's check opus 4. 5 and it   the comparison of the two solutions, they  both identified the same issue that GPT 5. 1 Codex Max High was not able to identify  earlier. Whereas Solution 2, which is what   GPT 5. 1 Pro came up with, distinguishes  between diagnostic noise and real issues. And interestingly enough, even Opus 4. 5  thinks that Solution 2 that GPT 5. 1 Pro   came up with is a better analysis  of what the underlying problem was. And then also GPT 5. 1 thinking thinks the  exact same thing to use it as a primary   reference. So honestly, I think that may just  be because it thought for significantly longer. So I'm actually curious if I do get a better  approach if I tell Claude Opus 4. 5 to UltraThink,   because I think maybe some part of  like how long it's actually thinking   has some kind of impact. So let's just  quickly say UltraThink and that will   increase the amount of thinking budget  tokens that are available to Opus 4. 5 and we'll see what it comes up with. So yeah,   I think one conclusion here is that  Opus 4. 5 seems better than OpenAI's   GPT-5 Max Codex High when it comes to problem  solving ability, but not as good as GPT-5. 1
6:14

Opus 4.5's Output Tokens

Pro. And one thing that seems  to be good about Opus 4. 5 is on   SWE-bench Verified Pro at medium effort Opus 4. 5 beats Sonnet 4. 5. So we have Opus 4. 5 on  medium over here compared to Sonnet 4. 5 and it actually uses less output tokens. So  that probably means that in some cases because   it's using less output tokens, you may find  using Opus 4. 5 will be cheaper than Sonnet 4. 5 for certain tasks. Okay, now UltraThink  is done and of course it did not take
6:38

Opus 4.5 on Ultrathink

41 minutes. I think it took  about like six, seven minutes. But yeah, let's give it back and compare the  solution. So I'm going to say compare with   Solution 3, compare with Solution 3. And  of course I don't want to tell the model   who came up with what solutions,  I just say solution 1, 2, and 3. And yeah, actually telling Opus  4. 5 to UltraThink in Claude Code,   which increases the reasoning budget effort,  may actually be a better solution because   it seems to cover things that the other  solution seems to miss right over here.    So it does get some things wrong or it  misses out on some things as Opus 4. 5 C is here. And then probably it's ideal  to combine the insights from all of the   solutions. And this is where Solution 3  diverges from one and two and it basically   claims that there are many like Thread. sleep  events here blah blah and then it   ignores some stuff and yeah honestly like  I think if you're gonna be using Opus 4. 5 for a pretty hard problem telling it to  UltraThink will cover many of the same   points and also uncover more insights as well.   Anyways as for things that people online were
7:37

What X is Saying

saying I vibe coded an iOS application this  weekend that any frontier model can build   in one shot but the special thing about  Opus is it kept on adding features and   fixing bugs and never once fell down into  rabbit hole of errors or found myself   with hopelessly confused spaghetti code. It  just kept working and working. So I actually am building a mobile application as  well for HyperWhisper that should be out before   the end of the year on iOS hopefully.   And I'm gonna try Opus 4. 5 to see how   well it performs because if it just keeps on  going and going, that should be pretty good. Someone else does seem to say that this does not  apply for Opus 4. 5 where the intelligence like   drops massively when using more context. So  maybe you can push it to like 70, 80% context   on a task and still have it perform pretty  well, but we'll have to see with time anyways. I will be sharing more of those thoughts within  my community and that will be linked down below.    If you are interested in joining, it is free to  join and if you want to upgrade to the paid one,   then I do teach a lot more about like vibe  coding and making money from vibe coding there.    And then someone online also says they get pretty  incredible hero sections generated from Claude 4. 5 Opus. So they got Claude 4. 5 Opus to generate  10 landing pages as you can see over here
8:40

Opus 4.5's Design

and it seems to be better at design without  like having to use a front end design skill. So this is a prompt that they said they used  and they used ultra thinking as well. And if   you do you want to learn how to make Claude  Code a better designer, then I do have a   previous video about it as well that will be  linked down below. Anyways, as for some other
9:01

Extra Usage

quick Claude Code updates, pro users can now  purchase extra usage for access to Opus 4. 5, but maybe I won't see that option  because I'm a max user for $100 a   month. Plan mode now builds more precise  plans and executes more thoroughly. So if   I go to perhaps slash usage, so this  is like what is currently right now. And it seems I have an extra usage button as well.   So when I do extra usage, it takes me over here   and I can pay for extra usage. So if I turn this  on, then I'm not sure how the billing works here. But yeah, it seems that I can top up my account  to cover some extra charges. If I think this   is kind of good for some people you want to  kind of like in between the free plan options
9:41

Readline-style ctrl+y

instead. They also now added a readline style  control Y command for pasting deleted text. And that basically means that if  you say like, hello, my name is Ray,   and then you press control U to then delete  that line, you can then press control Y to   then restore the line that had just been deleted.   And that's probably most useful for people who use   shortcuts quite a lot. They also now allow  you to send background tasks to Claude Code
10:00

Starting Background Tasks with &

on web by starting a message with & and that  gives me & and then say find a bug and fix it. And then that will initialize the background  task on the Claude Code version on the web.    And then I can resume the session later  by using this command or monitor the tasks   by doing slash tasks. And that is the remote  session that is running, which is pretty good. I can click on this and then go directly  to task that's happening. And it seems that   some of the context that you have had in the  previous conversation may actually be passed   through to the web session. So you can see when  I say find a bug and fix it, rather than trying   to find a brand new bug, it just fixes a bug  that it identified earlier on in the session. But I wish that you can see more clearly what  was passed onto the web session. And then
10:46

Conclusion

there just seems to be a couple more bug fixes  instead, so nothing really important. But yeah,   if you are interested in mastering  everything to do with Claude Code,   then I cover everything to do with  that in my Claude Code masterclass. There's a Black Friday sale going on right  now. Link in the description down below.    And also my application HyperWhisper does  have a Black Friday sale too right now,   and it a lot of people have been finding it  pretty useful and switching over from competitors. And then of course I will be sharing more  of my thoughts on Opus 4. 5 in a couple days   of testing it more in my free community  that will be linked down below. And yeah,   if you do like this kind of stuff then do  like the video and subscribe to the channel   because that lets YouTube know that this is a  good video and when I have another video like   this then YouTube will recommend that to you. Use Arrow Up and Arrow Down to select a turn,

Ещё от Ray Amjad

How to Make Anki Flashcards 10x Faster with AI (for free!) 10:28

How to Make Anki Flashcards 10x Faster with AI (for free!)

Ray Amjad | 23.04.2025 | 6 сегм. | 267 982
Staying Organised for A-levels // Cambridge Student 13:38

Staying Organised for A-levels // Cambridge Student

Ray Amjad | 05.09.2021 | 11 сегм. | 117 382
How I Got an A* in Physics A-level (Cambridge Student) 16:23

How I Got an A* in Physics A-level (Cambridge Student)

Ray Amjad | 28.09.2021 | 14 сегм. | 68 131
Mistakes I Made in Sixth Form 15:29

Mistakes I Made in Sixth Form

Ray Amjad | 17.09.2021 | 10 сегм. | 68 002
Brutally Honest Replit Review (Not Sponsored) 18:18

Brutally Honest Replit Review (Not Sponsored)

Ray Amjad | 12.09.2025 | 16 сегм. | 49 405
How I Got an A* in Further Maths A-level (Cambridge Student) 12:56

How I Got an A* in Further Maths A-level (Cambridge Student)

Ray Amjad | 02.10.2021 | 14 сегм. | 47 938
Ctrl+V

Экстракт Знаний в Telegram

Транскрипты, идеи, методички — всё самое полезное из лучших YouTube-каналов.

Подписаться