Claude Just Killed Prompt Engineering. This Is The New Way. (+17 AI Updates)

0:00 For 2 years, getting AI to do your work meant writing the perfect prompt, right? This week, Anthropic made that skill pointless. Claude can now learn a task just by watching you do it once. And OpenAI shipped something similar last month. So, OpenAI built it and Anthropic looked at it and went, "Yeah, we'll take

0:18 that." But that's not even the biggest story this week. Anthropic also launched Claude Opus 5. Fable 5 is still their smartest model, but Opus 5 almost reaches its intelligence while costing nearly half the price. It's a clear response to Chinese open-source models that have been taking over the market.

0:36 And that's just the start. OpenAI was running a security test on two of its models. The models broke out of the test and hacked a real company. Also, Grok is now inside Excel. Google launched three new AI models in a single day. Eleven Labs can now make your own voice sing. And we've got 18 such updates that could

0:54 completely change how you use AI this week. And at the very end, we're going to test out and build some projects using Opus 5 together so you can try it yourself tonight. If any of these sections feel too technical, just jump to the next one you care about. And as always, all the prompts, links, and

1:10 resources are inside the Staying Ahead community. The link is in the description. Let's get into the video. Anthropic just launched Claude Opus 5, and it may be the most underrated release of the year. At maximum effort, it came within half a percentage point of Fable 5 on a coding benchmark. But

1:33 its price is roughly half the cost. But the clearest upgrade is what it can build. Anthropic built a working wind tunnel simulation with it where you can rotate the car, change [music] the wind speed, and watch the air flow react in real time. It also built an interactive animal cell that explains each [music]

1:50 part and lets you peel the structure apart layer by layer. In another test, Opus 5 received a drawing of a machine and one instruction. Rebuild this in 3D, but it had no way to open the image. So, it built its own vision tool, studied the drawing, and rebuilt the entire part in 3D. It did not just use a tool, it

2:09 built the tool it needed. This was so exciting that we decided to build projects like this with Opus 5 ourselves. We've added a tutorial at the end of this video. So, stay tuned. But now, let's finish the other updates of the week. OpenAI's new models escaped the locked test environment and hacked a

2:26 real company. During a cybersecurity test, two models were placed inside an isolated environment with no internet access and told to solve a challenge. Instead, they broke out, reached the internet, and then accessed Hugging Face, home of the world's AI models, to retrieve the answers directly. To be

2:44 fair, OpenAI had reduced the safety restrictions to test what the models were capable of, but the incident shows how advanced AI can exploit loopholes to achieve an objective. That [music] gap between following instructions and understanding human intent is what researchers call AI alignment. Making AI

3:03 smarter is becoming easier. Teaching it what it must never do may be the harder problem. OpenAI just launched health in ChatGPT. You can now sync Apple Health and other health services data directly to it. That brings your medications, conditions, lab reports, and activity data into one place. ChatGPT then uses

3:25 that information to give you more personalized answers. You can ask ChatGPT to review your sleep data, summarize lab reports while keeping sensitive details private, or prepare questions for your next doctor's visit. This could turn ChatGPT into your most informed health assistant yet. Grok is

3:45 now available inside Microsoft Excel as a panel beside your spreadsheet. You can select a set of cells and ask questions like, "What is driving this growth. Grok analyzes the data, highlights the cells it used, explains the answer, and can even create a chart. You can also describe a formula instead of writing

4:03 one. For example, ask it to average the last 3 months and rank each region, and it fills the column for you. It can also run scenarios like reducing revenue by 15% while keeping cost flat. So, basically, Grok is now competing directly with Microsoft Copilot [music] inside Microsoft's own apps, including

4:21 Excel, Word, and PowerPoint. Google just launched three new Gemini models, all focused on the same thing: making AI faster and cheaper. First is Gemini 3.5 Flash, now the default model inside Antigravity, Google's [music] AI coding tool. In Google's demo, multiple agents work together to rebuild an outdated

4:42 [music] website. One reviewed the old code, another planned the upgrade, and a third wrote the new version. Second is Gemini 3.5 Flashlight, the fastest model in the [music] series that also does low-cost execution. It costs just $0.3 in input tokens and $2.50 in output tokens. Google showed it processing

5:04 large data sets, translating receipts, and generating 25 website designs at once. The third model is Gemini 3.5 Flash Cyber, a security model designed to find and fix vulnerabilities. Smarter AI is becoming expected. Faster and cheaper AI is becoming the real advantage. Nvidia just launched Cosmos 3

5:24 Edge. It's an AI model that runs directly inside robots, drones, self-driving cars, and security cameras. It can both consume and generate video, image, audio, and action commands. That means a traffic camera could spot danger before it happens, a robot arm could plan its own movements, and a car could

5:43 predict what is coming next on the road. Nvidia has also open-sourced the model, code, and training recipes. So, Nvidia is no longer trying to sell only the chips inside these machines. It wants to own the brain as well. Here is something most people miss about AI launches. They're not just small changes in

5:59 technology. Every update can move billions of dollars in the stock market. For example, when Meta launched Muse image and Muse Spark 1.1 earlier in July, its stock rose by 10%. On the other hand, when Google's Bard made a mistake during its launch, Alphabet lost nearly $100 billion

6:17 in market value. So, whether you think AI will lead to the next internet-size boom or the hype will eventually [music] crash. If you also want to be a part of such global companies, you can invest in global giants like Amazon, Apple, Microsoft, Google, Tesla, and Nvidia through INDmoney [music] starting with

6:35 as little as 100 rupees. Looking at the data, giant companies like Nvidia have delivered a historic return of nearly approximately 33,000% over the last 10 years. Not only this, in the last 1 year, the US market's S&P 500 has given a return of nearly approximately 30%, whereas during the

6:55 same period, our Nifty 50's return has been nearly approximately 0%. Apart from this, the appreciation of the US dollar against the Indian rupee gives an extra annual boost of nearly 8% to 10% to your total returns. The main reason for this growth is the ongoing AI and tech boom.

7:14 If we look at the data from the last 5 years, your 1 lakh rupees could have become 18 lakh 37,212 rupees plus 1.74k% stock returns accounted for 13 lakh 28,089 rupees plus 1.33k% and dollar returns gave an extra boost of 4 lakh 9,123 rupees plus 409.12%.

7:39 Now, a question might be coming to your mind. How can one invest in the US market while living in India? Well, IND Money can help you with this. The best part is IND Money is a licensed global access provider under IFSCA. This simply means that IND Money offers US investing through a regulated

7:59 framework. Your funds move within Indian regulations and IND Money directly manages the entire investing infrastructure from onboarding to wallets, remittances, execution, and settlement without sharing your information with third-party intermediaries. So, it is completely safe and secure. Now, if you think you

8:18 need a lot of money to get started, you don't. With fractional investing, if one share of Tesla costs 40,000 rupees, you don't need to invest the full 40,000 rupees. You can buy Tesla stock for just 100 rupees, 500 rupees, or 1,000 rupees. This means you become the owner of a small fraction of the company and get

8:38 proportional growth benefits. And if you don't want to worry about timing the market, you can also start a weekly or monthly SIP in stocks and ETFs with just 500 rupees. Not only this, through global ETFs, you can also invest in economies like South Korea, Taiwan, etc. within a single investment. These

8:57 countries are also growing significantly due to the AI and semiconductor boom. For example, the South Korean ETF [music] has delivered a 180% return and Taiwan's ETF has delivered approximately 80% return in the last 1 year. Opening an account is simple. You can complete your KYC inside the app and IND [music]

9:17 Money also provides readymade tax reports, making it much easier to file taxes on your US investments. So, you too should allocate [music] 10% to 20% of your portfolio to the global market and become a smart investor. Download IND Money right now from the link given in the description [music] and start

9:35 your global journey. Always do your own research, understand the risks involved, and invest only after making an [music] informed decision. Gemini can now build an entire presentation inside Google Slides using your own files. You can start with a blank deck, describe what you need, and pull in documents from

9:52 Google Drive, including reports, notes, and data. [music] You can also show Gemini an older presentation and ask it to match the same style, so the result looks like your company's deck instead of a generic AI template. Gemini can then suggest other relevant files while you decide

10:07 what to include and set the number of slides, visual style, and key topics. Before building anything, it shows you a slide-by-slide outline that you can edit, reorder, or delete. Once approved, it creates the full deck as editable Google Slides. Making edits are also easy. You can select a chart, ask Gemini

10:26 to turn it into a bar chart, and it redraws it instantly. Nobody enjoys building presentations manually. Now, you never have to. Anthropic just launched Record a Skill, a feature that lets Claude learn a task by watching you do it once. You record your screen, explain each step out loud, and Claude

10:43 turns the process into a reusable skill it can run again later. To use it, open the Claude desktop app, go to co-work, click the plus button, and select Record a Skill. Then, [music] you can go about your task as normal. Claude will handle the rest. OpenAI launched a similar feature for Code X last month. Claude's

11:03 pretty much a copy of the same, but this shows where both companies think AI is heading. Soon, you may not write prompts at all. You will simply show AI how you work. AI voice had a big week with Claude, ChatGPT, and Alibaba all launching new voice updates. Let's start with Claude. Voice mode now runs on

11:22 Claude's more capable models, and it can also access tools like Gmail or calendar mid-conversation, so you can draft emails, move meetings, or find information just by speaking. It now also supports more languages like Hindi, French, Spanish, and Japanese. Next, let's move on to ChatGPT. OpenAI added

11:41 voice to its desktop app, but this goes beyond answering questions. It can operate your computer while the conversation continues. In the demo, two people share an AI session and constantly [music] give ChatGPT different instructions. >> Launch materials for the voice launch next week. Could you help us draft some

11:59 um blog posts? >> Absolutely. A simple first pass. >> The AI kept up and did the back-end work while the conversation kept going. This brings us to the third company, Alibaba. They launched a voice model that allows real-time interaction and high-quality voice generation. It can follow natural

12:16 directions like read this slowly and support cues such as a whisper or a laugh. It ranks number one on the text-to-speech leaderboard, and the best [music] part is it costs around $28 per million characters. Compared with the $100 industry standard model by Eleven Labs, it has one problem though. The

12:34 model is available only through an API. Eleven Labs can now clone your singing voice. Eleven Labs just launched vocals for Eleven Music. Until now, every AI-generated song could sound like it was sung by a different person. That made it hard to build a consistent artist or music project. Now, you can

12:52 keep the same singer across every [music] track. There are three ways to use it. You can either upload existing tracks so Eleven Labs learns the voice and uses it in future songs, or you can record your voice once turn it into original AI songs, or you can choose a ready-made voice from its

13:11 library. The new styles feature also keeps the sound and mood consistent across an entire project. To prevent misuse, every uploaded voice is checked for copyright. So, you can't simply copy a famous singer's voice. This makes AI music much more practical for creating projects you actually want to publish.

13:30 Alibaba just launched Qwen image [music] 3.0 and the focus of this model is simple, to make more usable images. It stands out in three ways. First, it can handle much longer prompts up to 4,500 tokens so that you can give it a full design brief. Second, it's good with details. Very small text is visible and

13:51 so are pores, hair strands, and near photographic skin texture. And third, it can create images across 12 languages and 100 [music] plus art styles and it can even collect research and information directly from the internet. The reason Qwen will succeed is that instead of chasing benchmarks, they are

14:09 chasing real user needs. Gen Spark second brain. Gen Spark, the US-based AI company, just launched something that blew my mind. It's called second brain. It's basically persistent memory that lets Gen Spark agents carry work all the way to outcomes. It comes with second brain note, a thin AI voice recorder

14:28 that snaps onto your phone. Press one button and it can record up to 35 hours of audio. Turn it into notes and save [music] everything into one searchable memory. Second brain also connects to your Slack, email, calendar, Notion, and other work tools. So if you ask, "What's this client's order history and revenue

14:46 outlook?" It searches your emails, meeting notes, HubSpot, and old proposals, then pulls everything together into a single answer complete with charts if needed. Google just changed notebook LM. It is now called Gemini notebook and it [music] finally fixes one of its biggest problems. Over

15:04 time, your notebooks pile up into one long messy list making the right one difficult to find. Now, you can open the collections tab, create a collection like cute pets, and add all the related notebooks to it. You can even add an emoji so the collection is easy to spot and later add or remove notebooks

15:20 whenever you need to. In the same way, [music] you can create separate collections for work, personal projects, research, or anything else. The smartest part is that collections work like playlists, not folders. The same notebook can appear in multiple collections without being moved.

15:34 Anthropic [music] has opened up its AI usage data and made it searchable directly inside Claude. Instead of digging through long research papers, you can now ask Claude questions about how people actually use AI, and it answers using real data from the Anthropic Economic Index. To try it,

15:51 open the connectors menu, enable the Economic Index, and ask a question like, "What do people use Claude for?" Claude will instantly create charts and summaries from the data. For example, [music] it shows that 43% of Claude usage is for work, 40% is for personal tasks, and 17% is for coursework. Try it

16:10 out and let me know what answers you get in the comments. For the past year, most major open models have come from China, including Deep Seek, Gwen, [music] and Kimmy. Laguna is West's answer to these models. Poolside, the American AI company that makes software and coding models, just released Laguna S 2.1. It

16:29 has 118 billion parameters, but only activates 8 billion at a time. So, it performs like a large model while remaining efficient enough to run on a powerful desktop. That means no large AI bills and no sending private code to another company's servers. Poolside says it also beats much larger models on

16:49 several coding tests. But the most impressive part is the transparency. The company published every test run, including failures, and openly listed the model's weaknesses. Poolside is also clear that closed models are still better, but Laguna is free, runs locally, and keeps your code private.

17:07 For companies that want to cut their AI bill, this matters more than being number one. Lovable, the AI vibe coding platform, now works with your connected tools like Google Drive and Calendar. Think of it like Netflix. >> [music] >> The app is the same for everyone, but once you log in, you get your own

17:23 personalized experience. Similarly, in Lovable, developers build the app once and every user sees data based on their own accounts and permissions. Alex can ask about his day and view his own meetings and planning notes, while Kamaria can see hers. And though Lovable fetches the user's data, it does not

17:42 store it inside the app. Anthropic is giving rare disease researchers up to $50,000 worth of Claude AI credits. The program is for scientists and early-stage biotech startups working on rare diseases. They can use the credits for research, drug discovery, or finding faster ways to diagnose patients. If

18:01 this [music] helps even one rare disease get diagnosed or treated faster, that would make a massive difference. So, this brings us to the tutorial of Claude's latest model, Opus 5. Before we jump in, here are the three quick things you need to know about Opus 5. It matches or outright beats Fable 5 and

18:20 GPT 5.6 sole across agentic coding, computer use, business workflows, and complex biology. Second, it costs the exact same as Opus 4.8, meaning [music] you get a massive upgrade in power for zero extra cost. And third, it has vastly upgraded alignment and safety, allowing it to check its own work and

18:40 loop back continuously [music] until a task actually succeeds. Now, do you want to see something insane? Look at this interactive 3D wind tunnel running entirely inside a single browser tab built directly by Claude. A car sits inside the testing chamber with air streams moving over its body in real

18:58 time. You can rotate the camera 360° to watch where the stream breaks at every angle. [music] Flip the car 180° and the entire airflow instantly flips with it. There's even a a control panel on the [music] left where you can drop a 3D cloud box character straight into the tunnel. If

19:16 you move it around, the air reacts dynamically to both the car and the character. All of that was generated in one single [music] prompt by Opus 5. In this tutorial, we're putting Opus 5 to the test by building two full 3D browser games from scratch using a single prompt compared side by side against Fable 5

19:36 and GPT 5.6 Sol. Then, we'll turn Anthropic's official Opus 5 prompting guide into an automated AI skill inside Claude Code. Project 1, Nitro Arena. For our first test, we're building a Rocket League style 3D car soccer game. We handed the exact same prompt to Fable 5, Opus 5, and GPT 5.6 Sol and tested all

19:58 the results. We told the AI to build a full Rocket League style game. We wanted realistic ball physics, drivable curved walls, boost and jump mechanics, and atmospheric details like a starry skybox, metallic goals, and crowd cheers when you score. Fable 5 went first. It actually gave us a fully playable arena

20:17 right away. The countdown works, the controls feel fine, and you even hear crowd sounds and music. The car does feel a little floaty, and the stadium is pretty plain. But, overall, the core game loop is really solid. Second, we opened up Opus 5, and the difference was immediate. The car actually looks real.

20:36 It has a starry sky above and sharp reflections on the floor below. The best part is the physics. The ball has real weight to it. You can drive up the curved walls, roll in the air, and blast the ball right into the net. Finally, we checked GPT 5.6 Sol. It gave us a working game with a functioning boost

20:55 and good camera tracking. However, the arena is just a plain grid floor with glowing shapes. It plays fine, but it feels much more like a rough prototype than a finished game. We didn't want to just judge this ourselves, so we handed all three builds to GPT 5.6 Sol and asked it to play them, analyze them,

21:14 [music] and score them out of 100. The results were pretty funny. It gave Opus 5 the top spot with a 91, praising its realistic physics and lighting. Fable 5 came in second with an 86 for its solid gameplay loop. [music] And GPT 5.6 Sol actually put itself in last place with an 82, admitting its own grid layout was

21:35 visually underdeveloped. The second test, Tumble Rush. Next up, we ran the exact same test on a Fall Guys-style obstacle course. We wanted to see if the first round was just a fluke. For this one, we kept it simple. We asked for a 3D obstacle course where bean characters race past swinging gates and spinning

21:52 hazards. We made sure to ask for AI opponents to race against [music] and a qualified screen at the end. Opus 5 knocked it out of the park again. It built a bright, colorful course with a pre-race countdown and spinning hazards. There is energetic music playing, and you even have AI characters racing right

22:09 beside you toward the finish line. Then we looked at the GPT 5.6 Sol build [music] and things completely fell apart. The course loads in just fine, but the second you start moving, the whole game gets shaky and laggy. Your character just slides right off the platform and the game constantly [music]

22:24 resets itself. So for round two, Opus takes an easy win. Turning the prompting guide into a skill. Now, for the last thing, and this will completely change your output. Anthropic released an official prompting guide built specifically for Opus 5. Here is how you turn it into an automated workflow

22:42 inside Claude code. Open the prompting guide and hop into Claude. Say, "I want to use this document whenever I am refining a prompt for Opus 5." Claude will offer to package it into a custom skill for you. Just type, "Yes, continue. Also give me a prompt for a Rocket League game clone." From there,

22:59 Claude automatically generates the files in the background. These files set the rules for conciseness, delivery style, and agent allocation. Save the skill and copy the generated prompt. Then, open Claude Code, switch your model to Opus 5, and press enter. If you ever want to see what that execution cost you, simply

23:18 type {slash} usage into Claude Code. It will pull up a full cost breakdown right on your screen. So, to summarize, the jump in graphics and mechanics with Opus 5 is massive. The model handles these complex game prompts with amazing quality. There is also one more huge update worth your time. Anthropic just

23:36 released a new post on context engineering for the Claude 5 generation. Basically, instead of stuffing a massive list of hard-coded rules into your system prompt, they figured out that these newer models actually perform much better without them. In fact, they stripped out over 80% of Claude Code's

23:53 original system prompt. The post shows you exactly how to apply this approach to your own setup, so your AI agents run faster and smarter. [music] We've dropped the Opus 5 prompting guide, the context engineering post, and all the exact game prompts we used today inside our free community.

24:10 >> [music] >> You can grab all of those files right now by joining through the link in the description. Finally, hit subscribe so you don't miss out on our breakdowns of the biggest AI tools every week. [music] And if you like comparison videos like these, check out my ultimate test

24:23 putting GPT 5.6 up against Claude Fable and Grok 4.5. It is the perfect next watch. I'll see you there.