YouTube2Text

Claude Code'u 27 Kat Ucuza Kullandım — Transcript

by Selma Kocabıyık · 2,401 words · 365 segments · language en · Watch on YouTube

Full transcript

  1. 0:00A few weeks ago, a new model appeared on
  2. 0:02OpenRouter, and everyone started calling
  3. 0:03it Space Bunny; according to MiniMax's
  4. 0:05statements, it even held the top spot
  5. 0:07for 10 days in a row. Afterward, it
  6. 0:09became clear who owned this model.
  7. 0:11It turned out to be MiniMax's preview
  8. 0:13model, and the one we will be testing
  9. 0:15today is M 3.1 Flash, which is its
  10. 0:17official and improved version.
  11. 0:20As you know, I usually only talk about
  12. 0:22Claude Code on this channel, and I've
  13. 0:23barely even shown Codex. That's why
  14. 0:25we're going to do something different in
  15. 0:26this video, and I'm going to try out a
  16. 0:28different model.
  17. 0:29We'll take this new model from MiniMax
  18. 0:31and Sonnet 3.5, put them both into
  19. 0:32Claude Code, conduct three experiments,
  20. 0:34give them the same prompts in each, and
  21. 0:36examine the results together.
  22. 0:39I will show you in a table later how
  23. 0:40many tokens they spent and how much they
  24. 0:42cost, and we are conducting this
  25. 0:44experiment with the sponsorship of
  26. 0:46MiniMax.
  27. 0:47Now, you know that in our agent
  28. 0:49architecture, we normally assign a
  29. 0:50different model to each part of the
  30. 0:52architecture when building an agent team
  31. 0:54.
  32. 0:55Usually, in my agent company, the model
  33. 0:56handling routine tasks might be Sonnet,
  34. 0:58while the model doing high-level
  35. 1:00planning or oversight might be Opus.
  36. 1:02Or, for video production, which isn't a
  37. 1:04complex process, we use Opus 3.5 again.
  38. 1:07Considering the models currently on the
  39. 1:09market, one of the best ways to utilize
  40. 1:12this is by chaining different models
  41. 1:14together.
  42. 1:15What do I mean? Naturally, Claude Code
  43. 1:17is expensive, and its models are pricey
  44. 1:19too.
  45. 1:20But if we only use Claude's models for
  46. 1:22the necessary tasks and connect cheaper
  47. 1:25models for other tasks—for example,
  48. 1:27through a tool like OpenRouter or
  49. 1:29different API calls—then the agent
  50. 1:31architecture we've built will start
  51. 1:33working much more efficiently in terms
  52. 1:35of both tokens and cost.
  53. 1:38Given this situation, we actually need
  54. 1:40to try different models and figure out
  55. 1:42through trial and error which one fits
  56. 1:44our own workflow best.
  57. 1:47In this video, we will be testing this
  58. 1:49with MiniMax's sponsorship. Of course, I
  59. 1:51will show two models here, but I think
  60. 1:53you can conduct these types of
  61. 1:54experiments on other models available in
  62. 1:55the market as well. Now let's get
  63. 1:57started right away. Claude Code is a
  64. 1:58program; it gives commands to files, we
  65. 2:00run models inside it, it writes code,
  66. 2:02and so on.
  67. 2:03But you can actually use Claude Code
  68. 2:05with another model as well. We will keep
  69. 2:07the program the same and only change the
  70. 2:09model inside. Now, on the left, you can
  71. 2:11see Sonnet 3.5, and on the right, you
  72. 2:14can see MiniMax's M Flash model. Both
  73. 2:16are inside Claude Code now. We will give
  74. 2:18both the same prompt and start from an
  75. 2:20empty folder. The only thing changing is
  76. 2:22our model. So, how do you set this up?
  77. 2:24It works like this. At this point, I
  78. 2:25followed MiniMax's own documentation
  79. 2:27completely. There are two important
  80. 2:29points. Which address you connect to and
  81. 2:31which key you use. In other words, the
  82. 2:33endpoint and the API key are the two
  83. 2:34important factors here. The address here
  84. 2:36is the address compatible with MiniMax's
  85. 2:38Claude, and the key will be the key from
  86. 2:40your MiniMax account. Now, of course, I
  87. 2:42only provided this to the terminal on
  88. 2:44the right. This way, the window on the
  89. 2:46left remained as standard Claude Code.
  90. 2:48In the one on the right, we connected to
  91. 2:50MiniMax. Now let's check it. The status
  92. 2:52screen shows the address as Minimax, and
  93. 2:54the model line says M Flash. Okay, this
  94. 2:56part is great. Now, there is also
  95. 2:57something we call the reasoning level.
  96. 2:59In Flash, reasoning is always on and
  97. 3:00cannot be turned off. At this point, I
  98. 3:02set the level to the highest, Max, and
  99. 3:04to be fair, I set Sonnet on the left to
  100. 3:06Max as well. By the way, I wanted to
  101. 3:08provide a good prompt so that our final
  102. 3:09product, the output, would be good. I
  103. 3:11didn't want to test the models in just
  104. 3:12one direction. That's why we tested them
  105. 3:14with different scenarios in three
  106. 3:15different directions. In both, I asked
  107. 3:17for a 3D Mars scene that runs in the
  108. 3:19browser. You drive a rover, a dust storm
  109. 3:21starts, day turns into night, and so on
  110. 3:24.
  111. 3:25In the second one, I wanted a simulator
  112. 3:27showing the world 1000 years later.
  113. 3:30We wanted a 3D globe and a time bar to
  114. 3:32be included.
  115. 3:33I wanted it to show the changes in the
  116. 3:36world as you slide from 2026 to 3026.
  117. 3:40Our third direction was to have both of
  118. 3:42them create an AI interface that acts as
  119. 3:44if it were conscious.
  120. 3:46You type something, and the robot's
  121. 3:48thoughts and mood change live on the
  122. 3:50screen.
  123. 3:51I will check if the work is finished
  124. 3:53using the same list again. We will look
  125. 3:55together at whether it opens in the
  126. 3:57browser, if the interaction works, and
  127. 3:59if the visuals are smooth and fluent. I
  128. 4:01put it into the prompt screen. You can
  129. 4:02try the same thing on your own computer
  130. 4:04. We started both at the same time. I
  131. 4:06sped up the waiting parts, but I left
  132. 4:08the important moments for you.
  133. 4:10I would like to state one thing from the
  134. 4:11beginning. I turned off command
  135. 4:13execution permissions on both sides.
  136. 4:15Meaning both models could only write
  137. 4:16files to their own folders.
  138. 4:18They didn't touch anything else on my
  139. 4:20computer. And this made the test fair.
  140. 4:22Because both had to work with the same
  141. 4:24constraints. Now, I started both at the
  142. 4:27same time, and the first difference
  143. 4:28appeared in speed. The Flash model
  144. 4:30finished all three tasks between 8 and
  145. 4:3213 minutes. Sonnet 3.5, on the other
  146. 4:34hand, thought for much longer.
  147. 4:36It delivered the files between 38 and 48
  148. 4:38minutes.
  149. 4:40There is also something interesting here
  150. 4:41. Since neither model could execute
  151. 4:43commands, I couldn't test the code in
  152. 4:44the browser.
  153. 4:45The Flash model wrote this clearly. It
  154. 4:47said that since it couldn't run and
  155. 4:49verify the file, it checked the code
  156. 4:50manually and finished the job that way.
  157. 4:53Sonnet, however, called an assistant
  158. 4:55agent to check its own code, and that
  159. 4:57agent kept asking for permissions.
  160. 5:00I didn't grant those permissions and
  161. 5:02stopped it after one hour. So, Sonnet
  162. 5:04worked much more meticulously, but it
  163. 5:06was also quite slow. The Flash model
  164. 5:08approached the tasks more practically.
  165. 5:10Let's look at our Mars example first.
  166. 5:12This is the version Sonnet made. It set
  167. 5:14the scene at Jezero Crater. By the way,
  168. 5:16I'm hearing this name for the first time
  169. 5:17too. I had no idea. There is a small
  170. 5:19panel on the top right, by the way. It
  171. 5:20shows all the characteristics you can
  172. 5:22think of on Mars, like local time,
  173. 5:24temperature, wind, visibility, distance
  174. 5:26, dust density, and so on.
  175. 5:28Now, as you drive the rover, dust kicks
  176. 5:30up behind it. It added a nice effect
  177. 5:32like that.
  178. 5:33Uh, there is also a feature where we can
  179. 5:35turn the headlights on and off, by the
  180. 5:37way. You can also change the camera
  181. 5:39angles however you like. It ended up
  182. 5:41with quite a nice design like this.
  183. 5:43By the way, I think it did a pretty good
  184. 5:45job with the lighting and sky color. I
  185. 5:47mean, it looks realistic. Now let's see
  186. 5:49how M 3.1 Flash handled the same task.
  187. 5:53By the way, it named its vehicle RS7 as
  188. 5:55well.
  189. 5:56It placed eight measurement points
  190. 5:57throughout the map.
  191. 5:58You collect them as you go near them.
  192. 6:01So, it actually added a little game that
  193. 6:03I didn't ask for or specifically add to
  194. 6:05the prompt.
  195. 6:06There is a dust storm effect here too,
  196. 6:08and it even shows the storm's intensity
  197. 6:09and such in the top right. It also has
  198. 6:12three different camera angles, and it
  199. 6:14even added wind sound effects.
  200. 6:19By the way, the graphics are simpler and
  201. 6:21more angular, but I can say it actually
  202. 6:23thought about more things when you look
  203. 6:25closely. It added a few more features.
  204. 6:28It added its own personal touch.
  205. 6:30So, when we look at the Mars section in
  206. 6:32summary, both did a good job.
  207. 6:34Sonnet has better visuals, but we could
  208. 6:36say Flash was a bit more generous with
  209. 6:38the gameplay.
  210. 6:39I didn't see a clear winner, but I'm
  211. 6:41curious what you think. Write it in the
  212. 6:43comments. Let's move on to our second
  213. 6:45example right away. That is the World
  214. 6:46Simulator. Now, this is the version from
  215. 6:48Sonnet. As you can see, the continents
  216. 6:50are clearly visible. I'm showing you
  217. 6:52this now because I'll show the other
  218. 6:53model's version in a bit. You will
  219. 6:54understand what I mean.
  220. 6:57It displayed the lights, the bars, the
  221. 6:59temperature levels, sea levels, glaciers
  222. 7:01, and desert areas much more clearly.
  223. 7:05We can even see how Africa slowly
  224. 7:07yellows as time progresses.
  225. 7:09It also specified this in the top left
  226. 7:11corner. It wrote that this is a
  227. 7:13fictional scenario, not a prediction.
  228. 7:15I didn't ask for that, but it added it
  229. 7:17itself.
  230. 7:18You know that Claude's models are a bit
  231. 7:20more serious about these ethical issues
  232. 7:22.
  233. 7:23Now let's look at Flash's globe. The
  234. 7:25panel on the left is nice. It actually
  235. 7:26did a really good job, by the way.
  236. 7:28It shows the numbers and reflects the
  237. 7:30settings well, but the main thing we
  238. 7:32want to see—the Earth itself—is not
  239. 7:33fully visible.
  240. 7:35As you can see, it's covered in clouds.
  241. 7:37So, we can't fully see the continents. I
  242. 7:39mean, it looks more like Jupiter than
  243. 7:41Earth. I can't make out the continents.
  244. 7:43Of course, not wanting to give up until
  245. 7:45we got what we wanted, we decided to fix
  246. 7:47it and went back to the drawing board.
  247. 7:50This time, I didn't just tell Flash to
  248. 7:52fix it. I asked it to first read the
  249. 7:54code, find the source of the problem,
  250. 7:56and write a plan.
  251. 7:57Without changing the code, I had it
  252. 7:59execute the plan it wrote itself, and
  253. 8:00this is the result.
  254. 8:02Continents and oceans are clear. As time
  255. 8:04progresses, Africa yellows, the
  256. 8:05coastlines change, and the city lights
  257. 8:07turn on.
  258. 8:08Now I'm moving on to my third example.
  259. 8:10Our robot interface that has its own
  260. 8:13consciousness. Let's look at Sonnet's
  261. 8:15first. It gave the robot its own name.
  262. 8:18Very cute, by the way.
  263. 8:19Its face changes as you type something.
  264. 8:21On the right, you can also see seven
  265. 8:23emotion bars. Things like joy, curiosity
  266. 8:25, love, sadness, anger, fear, energy,
  267. 8:27etc. You see brain waves underneath, and
  268. 8:30an inner voice below that. So it's like
  269. 8:32the robot is thinking for itself. At the
  270. 8:34top of the screen, there's a section
  271. 8:36that says "thinking" and counts my
  272. 8:37memory as three messages. They've also
  273. 8:39added some specific questions at the
  274. 8:40bottom for the user to ask. So you can
  275. 8:43actually see how it performs without
  276. 8:45even typing anything. Now, let's take a
  277. 8:48quick look at the Flash robot.
  278. 8:50It's named it "Curiosity 7." It has mood
  279. 8:52and energy bars.
  280. 8:53It selects the most important word from
  281. 8:55the sentence you write.
  282. 8:56It also displays its intent, topic, and
  283. 8:59memory on the screen.
  284. 9:01And you can watch its stream of thought
  285. 9:03live at the bottom. I mean, it's simpler
  286. 9:06than Sonnet's, but it does more or less
  287. 9:08what we want it to do.
  288. 9:10By the way, not a single call goes
  289. 9:11outside in either of them, and
  290. 9:13everything runs within the page. It's
  291. 9:15already a system we built locally. Now,
  292. 9:17let's look at our results together.
  293. 9:20First off, Sonnet 3.5 ran for
  294. 9:22approximately 38 to 48 minutes.
  295. 9:24It burned through about 7 million tokens
  296. 9:27.
  297. 9:28I'm calculating the cost via the API,
  298. 9:30because since mine is included in a
  299. 9:31subscription, it doesn't actually cost
  300. 9:33me money this way, but the result is
  301. 9:35still the tokens spent.
  302. 9:38By the way, these numbers are for all
  303. 9:39three tasks combined. Now, let me look
  304. 9:41at the M 3.1 Flash. Let's say a maximum
  305. 9:43of 15 minutes per task.
  306. 9:45It spent 17 million tokens in total for
  307. 9:47all tasks, but the cost is quite cheap.
  308. 9:50Now, Flash spent more tokens, but since
  309. 9:52the tokens are cheap, the total for the
  310. 9:54three tasks is 14 dollars and 44 cents
  311. 9:56less than Sonnet.
  312. 9:57By the way, it actually got the world
  313. 9:59wrong at first, but the amount of tokens
  314. 10:02we spent to fix it was 9 million, and
  315. 10:04the cost for that part was separate.
  316. 10:06Also, while fixing the world, we first
  317. 10:08had Flash create a plan.
  318. 10:10And then we had it execute it. So, when
  319. 10:12you explain it step-by-step like that,
  320. 10:14it performs better. Of course, we don't
  321. 10:16need to take this single task as a
  322. 10:18benchmark, and results might differ for
  323. 10:19other tasks, but in general, the picture
  324. 10:21is very clear. So, here is my takeaway.
  325. 10:23Let's remember the architectural system
  326. 10:25I described at the beginning of the
  327. 10:26video. For example, you can use a model
  328. 10:28like this M 3.1 Flash for daily routine
  329. 10:30tasks. It's much cheaper than Sonnet. I
  330. 10:32, for example, use Sonnet in my agent
  331. 10:34architectures for these daily routine
  332. 10:36tasks.
  333. 10:37But here, for instance, the Flash model
  334. 10:38is quite cheap and works more or less
  335. 10:40the same way as Sonnet.
  336. 10:41Even though it spends more tokens, it's
  337. 10:43several times cheaper. This can
  338. 10:44definitely be added to your agent
  339. 10:47architecture. If you have a more complex
  340. 10:49task, then a different model can
  341. 10:50definitely be used. The whole purpose
  342. 10:52here is to try and reduce our costs as
  343. 10:54much as possible.
  344. 10:56For example, when I think about my own
  345. 10:58system, I can use expensive models for
  346. 11:00planning or very difficult tasks, while
  347. 11:02assigning routine or repetitive tasks
  348. 11:05that waste tokens to cheaper models.
  349. 11:08If you want to try the M 3.1 Flash in
  350. 11:10Cloud Code, you can find the link in the
  351. 11:11description and the pinned comment.
  352. 11:14You can also run it on your own computer
  353. 11:16with the exact same setup I used. And if
  354. 11:18you wish, you can also use this model in
  355. 11:20Minimax's own interface. It even has a
  356. 11:23separate implementation within Minimax
  357. 11:26Code.
  358. 11:27But of course, if you are already using
  359. 11:29Cloud Code and want to switch to a
  360. 11:30different model, you can just stay in
  361. 11:32the same interface and change the model
  362. 11:33.
  363. 11:34The decision is entirely up to you and
  364. 11:36your projects. Which models do you use
  365. 11:38the most? Let's meet in the comments.

About this transcript

This page contains the full transcript of Claude Code'u 27 Kat Ucuza Kullandım by Selma Kocabıyık, generated from the public captions YouTube serves with the video. The transcript has 2,401 words across 365 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

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

Free YouTube transcript tool

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