Prática de Shadowing: Codex Is Winning… So Why Am I Still Paying for Claude? - Aprenda a falar inglês com vídeo

The last two weeks in the world of AI agents have been very interesting.

Mark Zuckerberg and Meta released their new coding agent, MuseCode, to rival ClaudeCode and Codex.

And you can use MuseCode in the terminal to build apps or as a general agent.

And it only takes two minutes to set up.

I'll show you how.

But we have way more to talk about, like DeepSeat, Kimmy, and Quen, and more models coming out of China.

We also need to talk about Codex latest updates.

We also need to talk about Cursor

and how they're shifting from a developer tool to a super app just like Claude and Codex.

We will also discuss Google and how they might actually be out of the AI race.

And we need to talk about Buzz for creating teams of agents and all of the new updates there.

This is an agent-native update where we cover the most important news

and updates on Frontier AI platforms and models so you can use AI agents to be more productive.

I'm Riley Brown.

Let's go.

The first agent-native update comes from Meta, which is very rare.

They don't usually do that many things on the frontier.

But here we go.

Mark Zuckerberg on Twitter said, releasing MuseCode in beta today.

It's a terminal coding agent that takes on complete software engineering tasks across large repos,

planning changes, writing code, validating the results powered by MuseSpark 1.2, the model, a coding-focused model update.

And based on the benchmarks that he posted, it's right in between Opus and GPT 5.6 Terra, which is their middle model between Sol and Luna.

You can see Muse Spark 1.2.

So here are the prices of the meta model compared to 5.6 Sol and Claude Opus.

And if you add up the input cost and the output cost, you get $35 input plus output for GPT 5.6

and for opus you get 30 dollars

and for meta muse spark with the top model on there It's only five dollars and fifty cents.

So that's a total difference of like five to six x

So it's significantly cheaper than using the top models at open AI

Now when you start comparing it to open AI's GPT 5.6 Terra it gets pretty comparable

Which we'll talk about later because open AI is lowering the prices of their other model But always remember,

every single update that I talk about is best understood by actually testing it out, not just looking at some charts.

So now, how do we actually use MuseCode?

It's very simple.

What you should do is you should go to either ClaudeCode or Codex.

If you watch my videos, you know that I use Codex a lot more.

And I'm just going to go to Codex and say, hello there, Codex.

Can you please download the latest MuseCode by Meta?

Mark Zuckerberg announced it yesterday.

the terminal coding agent, please download it now so I can use it in my terminal.

So all we have to do is run this

and we will wait a few seconds and it's going to download it to our computer.

Okay, so now it's done.

And remember, this is a terminal coding agent, so you can run this in the terminal.

I can either go to the terminal app on my computer, or if I'm using the Codex app or ChatGPT on the Codex version,

I can press Command-J and this is the terminal.

And now I can type in Muse.

And now I need to select Trust and Continue or Quit.

I'm going to click Trust and Continue.

And here we can either log in with browser or set an API key.

And so we're logging into Meta Platform.

And so if I click Login with browser, it'll automatically take me to the browser and here I can actually just sign in with Meta.

Once you sign in, it'll look a lot like this, and now we can just use MuseCode.

So I can say, hello, what model are you?

And here it says, I am MuseCode powered by Meta Muse Spark.

And in order to use this going forward, I normally just do it inside the terminal, the normal terminal, and we can zoom in, and remember all you do is type Muse,

and now you're using it inside the terminal and you can full screen it if you want.

I'm going to say, hey, I want you to create a bowling simulator game.

Like it should be like Wii bowling.

I want you to make a Wii bowling and then I want you to run it locally so I can play it.

Now we're creating Wii bowling.

And I highly recommend just testing this out for yourself.

Do it for knowledge work tasks, get it to create documents, spreadsheets, that type of thing.

And this model is really fast.

And in my opinion, it's somewhere between Opus and Sonnet.

By default, it will tell you something like this sandbox blocks persistent servers.

So we actually need this to be in YOLO mode to

get it to automatically be able to run terminal commands so that it can run whatever app that you create.

And we can do this by going to Codex.

And we can open Codex and we can say something like this.

I need you to set it so that Muse code always defaults to YOLO mode.

Please just do that right now for every session.

And this will allow your MuseCode to do anything you want or do anything on your computer.

So you won't have to like give weird permissions.

And then after you do that, if you were to open up a new terminal session by pressing Command N,

and you were to type in MuseCode, here you can see you are on YOLO mode.

So now MuseCode has full control over everything and you can get it to do everything.

And so now I want to check on that bowling game.

And okay, so this is the game that it created.

We have bowling and we can up the spin.

So this will spin it left, which means I need to aim it like here.

Oh.

Okay.

Maybe I won't do spin.

There you go.

So we're using Muse Code.

This isn't the greatest game ever created, but that's probably a problem of prompting.

I just wanted to empower you with the ability to test it.

It's really easy to test out.

I highly recommend testing out the new meta terminal coding agent.

Okay for the second agent native update I want to talk about the changes to codecs over the past two weeks.

We're going to come to the mobile app in just a second.

Let's start off with the desktop app.

In the desktop app we have a few new changes.

In the side panel over here you'll see

that we have the normal view right here where i can see all of my chats

but there's also this notifications bar and

so this is kind of the activity pane

so instead of them showing in this like fixed project view right here it'll show you the most recent ones

100

and it's really intuitive to use and it'll show you the exact folder

101

that they're running in as well

102

so you can see where the directory is on your computer

103

so this is just kind of feels a lot more like notifications

104

and it shows the most recent one that completed which I personally use most of the time.

105

Okay for the second update to the desktop app

106

which is my favorite design update

107

that they've made in a very long time involves the in-app browser

108

and so I in this chat created a website and

109

so this is a website it's kind of ugly but that's okay

110

and if we were to full screen it so that the agent chat disappears.

111

Now they updated this bottom part right here.

112

So I can fully, I can click on this

113

and it opens up and I can very easily say like the line that goes beneath episodes.

114

I don't really like that very much.

115

Can you also please make the like dark purple behind the logo at the top better?

116

The whole top bar is pretty ugly.

117

Can you please fix that?

118

And so I can just run it right here

119

and it's really easy to see what the agent is doing and then I can pin it to the bottom.

120

So you can see here if I scroll down to the bottom, I see this little ChatGPT logo and now I can open it up and I can get rid of it.

121

So it allows me to fully immerse myself in the website, right?

122

This is basically full screen and if I ever want to make changes, I just come down to the bottom, press this OpenAI thing right here.

123

Now, what I do want to let you know is if you use Whisperflow, it will get in the way of this.

124

So as you can see here, I moved Whisperflow to the right side of my screen.

125

And if you have Whisperflow, see how it gets in the way?

126

So what you need to do is you need to come over Whisperflow

127

and just drag it to the left or to the right.

128

And I choose to put it on the right.

129

So now I can use Whisperflow.

130

Whisperflow is right here.

131

And the OpenAI edit is here.

132

And you can notice here that browser use is coming into play and it's actually controlling the browser.

133

So that's just a fun new design of the in-app browser inside Codex.

134

While staying on the topic of browsers, they also made updates to their Chrome extension.

135

And you can find this in plugins.

136

And if you just type in Chrome and you come here and you download this, when you go to your Chrome browser,

137

you will see this ChatGPT icon and you can move it to the front slot if you really like it.

138

And now what we can do is we can select this, and we can just say,

139

please prepare a tweet based on the recent memories that you have of me based on what we did today.

140

Come up with the best possible tweet.

141

And I can fire this off, control this browser, and make the tweet, but don't post it, just queue it up.

142

And so all of the chats that you do inside this Chrome extension sync with the Codex app.

143

So you can see this is called draft tweet from today.

144

I can very easily just go to chat GPT.

145

If we go to chat GPT, you can see that it says draft tweet from today.

146

And so we are connected to Chrome.

147

And so here it actually queued it up.

148

Okay.

149

So the biggest AI shift isn't better answers.

150

It's agency.

151

Okay.

152

Well, we need to work on the tweet, but it queued it up.

153

And if you want to open it straight inside Codex from here, all you need to do is go here you're going to hit these three dots

154

and click open in app and you're going to open chat gpt

155

and that will automatically open it directly inside the app

156

so you can basically just use the codex app inside chrome

157

and this is just a nice new update okay for the

158

last update involving chat gpt we have a new toggle here at the top

159

so now you can toggle between chat and work i just released

160

or i might be just about to release an hour-long video on chat gpt work it is the most in-depth guide ever.

161

I was wrong about Codex.

162

I said that I wish they didn't split up Codex and GPT work.

163

I was wrong.

164

GPT work is insanely useful.

165

It's completely cloud-based and I use it every day.

166

So I highly recommend starting to use this from your phone.

167

You can basically control your email calendar.

168

All of the plugins that you would find inside the chat GPT app,

169

right up here, you can basically fully control with GPT work.

170

So all of the apps you add, you can fully control from your phone, which is just incredibly useful.

171

And then after this update, it's just more accessible here at the top between chat and work.

172

I hardly ever use chat anymore.

173

I just use work.

174

For the next agent native update, I want to talk about this tweet.

175

ChatGPT desktop app, or formerly known as Codex, their biggest competitor right now is Cursor, not Claw Desktop.

176

And the reason I tweeted this is I have it on good authority.

177

I've talked to some people and Cursor is in a middle of a revamp.

178

They are about to make major changes to their platform

179

and it is going to become a full super app just like Codex.

180

And the first signs that we're seeing of this is you can now connect Cursor to Google Workspace.

181

This is their CTO at Cursor.

182

Everyone thinks of Cursor as a tool for coding.

183

We thought so too.

184

But inside the company, many of our use cases aren't coding at all.

185

Research, data analysts, bug triage, and project management, just to name a few.

186

As it turns out, coding agents are a pretty good foundation for all kinds of work.

187

Could be a sign of what's to come.

188

So very clearly, they're going to make changes to their platform.

189

Inside Cursor, you will see this Customize tab.

190

This is equivalent to the Plugins tab inside Codex.

191

And if you come to the top and you just type in Google Drive, you will see this Google Drive setup here.

192

I've already set it up, so you will need to authenticate it.

193

But once you authenticate it, which just means sign into your Google Drive files, you can then try it in chat.

194

Here, I have Cursor set up to the new DeepSeq V4 model.

195

And so now this model is basically free to use.

196

I'm not even kidding.

197

And so I'm going to go ahead and stop this and I'm going to say,

198

hey, can you please create a quick spreadsheet on just the steps on how to train an AI model, along with the description on how to do it,

199

make it look professional, and then send me the link to the Google Sheets for this.

200

So now I'm using DeepSeek V4 Flash, the cheapest model in the world by DeepSeek.

201

It's like basically free to do this, maybe a cent or two.

202

And it will create a Google Sheets link.

203

And now it's working inside cursor.

204

And there you go.

205

It says done.

206

Your Google Sheet is created and ready.

207

It worked for 44 seconds.

208

So it's really fast.

209

I can open this up and boom, here we go.

210

The point is, is it can control Google Docs, Google Sheets, Google Slides, anything.

211

And you can use a better model than Deep Seek V4 if you want, but that's just the one that I've been using for fun.

212

And for the fourth Agent Native update, we have three new models coming out of China.

213

We have Kimi K3, we have Deep Seek V4 Flash, and we also have Quen 3.8 Max.

214

And these models are all good in their own way.

215

Kimi K3 is the best out of all three of the new models, but it's also the most expensive.

216

I think many people were really excited by how good the model was, but they were underwhelmed by the price.

217

It is quite expensive.

218

It's like almost as expensive as Sonnet for certain tasks, but I highly recommend trying it out.

219

Then, DeepSeek V4 Flash is probably the most interesting,

220

because for certain tasks, it is 105 times more cheap than Fable 5.

221

And don't worry, I'll show you how to get this model set up inside

222

Cursor in just a second as well as all the other models.

223

But I do want you to keep in mind that regarding this DeepSeek v4 Flash model, which is really good for how expensive it is or how cheap it is.

224

They said this, DeepSeek said this, they said, we plan to raise the overall pricing for DeepSeek API services in the near future with a significant increase expected.

225

Please plan your usage accordingly.

226

So that's not good.

227

Many people are like, oh no, how expensive is it going to be?

228

However, Dax from OpenCode said on the upcoming DeepSeek price increase, we've been able to reproduce their current prices even on rented GPUs.

229

So this likely isn't because they're losing money.

230

It's traffic shaping because they're overloaded.

231

So he's saying

232

that we're going to be able to host this DeepSeek model in the US for a similar price that it is now.

233

They're saying that the price increase is just because so many people around the globe are trying to use it.

234

Klein posted this about DeepSeek v4 Flash.

235

While DeepSeek v4 Flash is significantly cheaper on price per token,

236

this could be misleading if the overall cost per task ends up being higher due to more turns being made.

237

However, Artificial Analysis reports DeepSeek completing the same benchmark tasks as Fable at 105 times lower cost,

238

which is absolutely insane.

239

So my take on this is as follows.

240

I believe that DeepSeek v4 Flash may get more expensive in the future.

241

Maybe it'll get three times as expensive.

242

Maybe it'll get five times as expensive in the short run

243

because there's so much traffic and so much demand for this model.

244

But I think over the long run, like over the next two months, this model will get significantly cheaper

245

or a model just as good as it will get as cheap as the model is now.

246

And so for the next week or two, I highly recommend trying these models.

247

The truth of the matter is, every few weeks, new open models are released, and the costs are simply going down.

248

And so, you can do so much with a model just as simple as V4 Flash.

249

A lot of tasks you can do, and this model is 20 to 100 times cheaper, depending on the task, than Fable.

250

So, I highly recommend testing out these models and seeing what you can do with them.

251

For the next Agent Native update, I want to talk about Anthropic.

252

And there's a vibe going around Twitter and even on YouTube that people are getting a little fed up with Anthropic.

253

Not only is Fable only on API usage now, so it's incredibly expensive to use Fable,

254

many people are showing disappointment surrounding their new models, Opus 5 and Sonnet 5, and that they completely wasted time building them.

255

And I'm not kidding.

256

A lot of people, you know, AI researchers, are literally saying that Opus 5 sucks.

257

John Ennis said that, But I have decided Opus 5, Extra High, is basically trash.

258

It doesn't use its thinking budget to do anything more productive.

259

It just uses it to thrash around and do pointless and sometimes harmful things.

260

Fable is a good model and I will miss it, but I think I'm finally ready to just let Claude go.

261

Or at least until they release a new model.

262

And I think finally more people are realizing that the vibe is shifting from Anthropic to Codex.

263

Again, at least for now.

264

Kamal Hussein said, It's crazy how the consensus shifted away from Claude being the favorite to Codex.

265

It's not just Fibes, things favoring Codex.

266

It's a better harness.

267

The Codex desktop is significantly better.

268

Better pricing, especially with the latest pricing updates to the new Codex models.

269

Less refusals, and you can use your subscription freely wherever you want, which is really cool.

270

Anthropic has a lot more guards against using your subscription on other tools.

271

However, I just tweeted this.

272

However, I just tweeted this prediction.

273

Anthropic is pulling back the slingshot and will try to go on another run soon.

274

I think they got high on their momentum from Q1

275

and Q2 and tried to launch too many things and their products got confusing and their momentum wore off.

276

And they just launched so many different products that people couldn't even keep track of everything.

277

Like when you tried to use Claude Cowork and connect it to your phone, it was called dispatch.

278

But if you did the same thing with Claude Code, it was called Claude Remote.

279

They released a Law product, they released Claude Design, which is a good product.

280

It just got buried under all the different product announcements.

281

They had an extension, and then they had the whole Mythos, Fiasco, Fable, Sonnet, Opus.

282

And honestly, their model releases, besides Fable, just haven't been that good since Claude 4.6.

283

All of them have felt relatively similar.

284

However, there's one thing that just keeps me using Anthropix products, and that is the fact that it is just so much better at frontend.

285

I gave Codex and Anthropix the same prompt, and this is what Codex created,

286

and it's just so much worse than Anthropix.

287

It is just so much better at frontend.

288

And this is not just like frontend for landing pages and other things like that.

289

It's also for spreadsheets, docs, and presentations.

290

Opus and Fable blows the OpenAI models out of the water for these like knowledge work type documents and front-end design.

291

For everything else, I think GPT 5.6 Soul is basically as good as Fable.

292

It's just so much better at front-end design

293

and it created this in one prompt and I just think it looks so good.

294

Like this is, I'm working on my website for Agent Native and it's just so much better.

295

So this is the one thing that keeps me using the Claw Desktop app.

296

It's just better at design.

297

And don't forget that the Claw Desktop app, if you go to home, there is design mode.

298

So you can use Claw Design directly inside the desktop app.

299

It's just they've launched so many things that so many people forget about Claw Design, and it's still really good.

300

And the final thing that I want to discuss is I believe

301

that we are entering a new era in the world of agents.

302

I think the first half of the year was kind of the open claw personal agent movement.

303

And I believe that we're moving into the team of AI agent movement.

304

Now, I've already made a video talking about Buzz.

305

And so this is Jack Dorsey's new platform where it is a Slack clone, basically, except it's made to be used with AI agents.

306

These are my existing codecs and clawed code running in this Slack-like interface

307

and I can at mention Claude and I can at mention Codex and say,

308

hey, work together to build an app that lets me use DeepSeq.

309

And you can at mention them and you'll notice here that they both reacted to it.

310

You can see Codex and Claude Code reacted to it.

311

And now you can see that they're both working on a response.

312

CloudCode and Codex are going to respond.

313

And I can message them as if they were just humans.

314

And they will actually work together to get it done.

315

And if you come to the agents tab, you can see your full team of AI agents.

316

And you can add Cursor.

317

You can add Devin.

318

And you can make any model the default model.

319

So I could come here and create an agent.

320

And I could customize the agent and I could use cursor for example

321

and I could choose any model from cursor

322

and for this I'm actually going to go ahead

323

and make the default model Kimmy K3

324

so now we're using Kimmy K3 with cursor

325

and I can give it instructions you are a content agent

326

and I can very easily name this agent cursor with Kimmy

327

and I can create an agent and now this agent is running

328

And so I could very easily go into my content channel and I could add, or I could actually just go like this.

329

Hey, Codex, please add Cursor with Kimmy to this channel and all other channels.

330

And Codex or any other agent can fully control this version of Slack.

331

And so it can add any agent to any channel, and it can even create agents.

332

And here at the bottom, you can see who's working.

333

So I can see that Codex is working.

334

And you can see that CursorWithKimmy is requesting approval.

335

And this is just kind of a team of agents that can do things together.

336

And you can see here CW, which is CursorWithKimmy, just responded.

337

And it said, I handled it myself.

338

I joined all seven channels, content, Vishal, coding, general research management, and notes.

339

And the reason I'm showing you this is I believe we're

340

in the very early innings of working with the team of AI agents.

341

It's not super easy yet.

342

The same way that OpenClaw wasn't very easy early on, and now GPT work is basically OpenClaw running inside chat GPT, and it can basically do anything for you.

343

And so I think we've advanced really far on the personal agent side, which is advancing these AI agents that we can access through our phones.

344

I think we're about to see the AI agent for teams really take shape over the next four to six months.

345

So definitely be on the lookout.

346

And as this progresses, I highly recommend

347

if you work at a large company to be the person

348

who can build a team of agents for your team

349

or create an agent that everyone on your team can access to get things done.

350

And I think we're going to see so many platforms like this.

351

We're going to see a lot more Slack agents that make agents really easy to configure.

352

And so that's something I would keep a close eye on.

353

Anyway, that's an update.

354

Those are the things that have really interested me over the past two weeks.

355

I really hope you like this video.

356

And if you could like and subscribe, it would help me out a ton.

357

I really appreciate you guys.

358

I will see you here for the next video.

✨ Vídeo recomendado

286 frases

Por que praticar a fala com esse vídeo?

O vídeo traz uma conversa dinâmica sobre atualizações em IA, repleta de termos técnicos e expressões coloquiais — o tipo de contexto perfeito para quem quer melhorar a fluência em situações reais. Ao praticar com ele, você aprende a articular ideias complexas de forma natural, além de se acostumar com a velocidade e a entonação de uma fala informal. É ideal para o shadowing em inglês: basta repetir frases logo após o falante, imitando o ritmo e a pronúncia. Essa técnica, aliada a um bom shadowing site, ajuda a treinar a audição e a fala ao mesmo tempo, tornando seu aprendizado mais eficaz.

Gramática e expressões no contexto

O vídeo usa estruturas práticas que valem a pena destacar:

"The last two weeks in the world of AI agents have been very interesting": A forma "have been" (present perfect continuous) é usada para falar de ações que começaram no passado e continuam no presente. Útil para descrever períodos recentes.
"It only takes two minutes to set up": A estrutura "it takes + tempo + to + verbo" é essencial para falar de duração de tarefas. Muito comum em conversas cotidianas.
"We also need to talk about Buzz for creating teams of agents": A preposição "for" aqui indica finalidade (para criar). Aprender a usar "for" corretamente evita erros com "to" ou "of".

Armadilhas comuns de pronúncia

Algumas palavras no vídeo podem ser desafiadoras. Veja como evitá-las:

"Codex": A pronúncia correta é /ˈkoʊdɛks/ (não "kódex" com som de "ó"). O "x" no final soa como "ks", igual em "text".
"Benchmarks": Fique atento ao "ch" no meio, que soa como "k" (/ˈbɛntʃmɑːrks/). Muitos aprendizes confundem com o som de "ch" em "chair".
"Terminal": A sílaba tônica é a segunda: /ˈtɜːrmɪnl/ (não "termiNal" com ênfase no final). O "i" soa como "i" em "sit".

Para dominar essas nuances, use o shadow speech: grave sua voz repetindo as frases e compare com o vídeo. Ferramentas como o shadowspeaks podem ajudar a identificar diferenças. Lembre-se: a chave para melhorar a pronúncia em inglês é a prática constante e a imitação de falantes nativos.

Gramática neste vídeo

As estruturas que o falante mais usa, com as palavras exatas do vídeo:

Estrutura	No vídeo
Frases condicionais if + oração, will/would + verbo — uma condição e seu resultado	if I click Login with browser, it'll automatically · if you use Whisperflow, it will get
Present perfect have/has + particípio passado — uma ação passada que ainda importa agora	have been · they've made · I've talked
Voz passiva be + particípio passado — o foco está no que acontece, não em quem faz	is run · is called · are connected
Orações relativas who / which + oração — informação extra sobre uma pessoa ou coisa	Meta, which is · Terra, which is · left, which means

O que é a Técnica de Shadowing?

Shadowing é uma técnica de aprendizado de idiomas com base científica, originalmente desenvolvida para o treinamento de intérpretes profissionais. O método é simples, mas poderoso: você ouve áudio em inglês nativo e repete imediatamente em voz alta — como uma sombra seguindo o falante com 1-2 segundos de atraso. Pesquisas mostram melhora significativa na precisão da pronúncia, entonação, ritmo, sons conectados, compreensão auditiva e fluência na fala.

Técnica de shadowing: leia o guia completo passo a passo →

Prática de Shadowing: Codex Is Winning… So Why Am I Still Paying for Claude? - Aprenda a falar inglês com vídeo

I Tested OpenAI's Dots vs. Meta's Muse. What You Need to Know.

Claude Code: The Advanced Guide (99% of Devs Skip These Features)

How Apple Accidentally Won The Al Race

5 open source tools that replaced my $320/mo AI stack...

Meta's new model wants "deep access" to your personal life...

Most devs don't understand what agents are

New Skills! v1.2 brings /wait-what, /writing-for-agents, and fixes /grill-me

How To Use Claude Better Than 99% Of People

Microsoft Surface vs. MacBook Air (Lab Tested)

Ultimate Guide To ChatGPT Codex for Everyday People

Fable 5 Took Over Unity and Built a Game in a Day

How Sam Altman Actually Uses AI to Run OpenAI and His Life

Por que praticar a fala com esse vídeo?

Gramática e expressões no contexto

Armadilhas comuns de pronúncia

Gramática neste vídeo

O que é a Técnica de Shadowing?