Claude Code'u 27 Kat Ucuza Kullandım — Transcript
Full transcript
- 0:00A few weeks ago, a new model appeared on
- 0:02OpenRouter, and everyone started calling
- 0:03it Space Bunny; according to MiniMax's
- 0:05statements, it even held the top spot
- 0:07for 10 days in a row. Afterward, it
- 0:09became clear who owned this model.
- 0:11It turned out to be MiniMax's preview
- 0:13model, and the one we will be testing
- 0:15today is M 3.1 Flash, which is its
- 0:17official and improved version.
- 0:20As you know, I usually only talk about
- 0:22Claude Code on this channel, and I've
- 0:23barely even shown Codex. That's why
- 0:25we're going to do something different in
- 0:26this video, and I'm going to try out a
- 0:28different model.
- 0:29We'll take this new model from MiniMax
- 0:31and Sonnet 3.5, put them both into
- 0:32Claude Code, conduct three experiments,
- 0:34give them the same prompts in each, and
- 0:36examine the results together.
- 0:39I will show you in a table later how
- 0:40many tokens they spent and how much they
- 0:42cost, and we are conducting this
- 0:44experiment with the sponsorship of
- 0:46MiniMax.
- 0:47Now, you know that in our agent
- 0:49architecture, we normally assign a
- 0:50different model to each part of the
- 0:52architecture when building an agent team
- 0:54.
- 0:55Usually, in my agent company, the model
- 0:56handling routine tasks might be Sonnet,
- 0:58while the model doing high-level
- 1:00planning or oversight might be Opus.
- 1:02Or, for video production, which isn't a
- 1:04complex process, we use Opus 3.5 again.
- 1:07Considering the models currently on the
- 1:09market, one of the best ways to utilize
- 1:12this is by chaining different models
- 1:14together.
- 1:15What do I mean? Naturally, Claude Code
- 1:17is expensive, and its models are pricey
- 1:19too.
- 1:20But if we only use Claude's models for
- 1:22the necessary tasks and connect cheaper
- 1:25models for other tasks—for example,
- 1:27through a tool like OpenRouter or
- 1:29different API calls—then the agent
- 1:31architecture we've built will start
- 1:33working much more efficiently in terms
- 1:35of both tokens and cost.
- 1:38Given this situation, we actually need
- 1:40to try different models and figure out
- 1:42through trial and error which one fits
- 1:44our own workflow best.
- 1:47In this video, we will be testing this
- 1:49with MiniMax's sponsorship. Of course, I
- 1:51will show two models here, but I think
- 1:53you can conduct these types of
- 1:54experiments on other models available in
- 1:55the market as well. Now let's get
- 1:57started right away. Claude Code is a
- 1:58program; it gives commands to files, we
- 2:00run models inside it, it writes code,
- 2:02and so on.
- 2:03But you can actually use Claude Code
- 2:05with another model as well. We will keep
- 2:07the program the same and only change the
- 2:09model inside. Now, on the left, you can
- 2:11see Sonnet 3.5, and on the right, you
- 2:14can see MiniMax's M Flash model. Both
- 2:16are inside Claude Code now. We will give
- 2:18both the same prompt and start from an
- 2:20empty folder. The only thing changing is
- 2:22our model. So, how do you set this up?
- 2:24It works like this. At this point, I
- 2:25followed MiniMax's own documentation
- 2:27completely. There are two important
- 2:29points. Which address you connect to and
- 2:31which key you use. In other words, the
- 2:33endpoint and the API key are the two
- 2:34important factors here. The address here
- 2:36is the address compatible with MiniMax's
- 2:38Claude, and the key will be the key from
- 2:40your MiniMax account. Now, of course, I
- 2:42only provided this to the terminal on
- 2:44the right. This way, the window on the
- 2:46left remained as standard Claude Code.
- 2:48In the one on the right, we connected to
- 2:50MiniMax. Now let's check it. The status
- 2:52screen shows the address as Minimax, and
- 2:54the model line says M Flash. Okay, this
- 2:56part is great. Now, there is also
- 2:57something we call the reasoning level.
- 2:59In Flash, reasoning is always on and
- 3:00cannot be turned off. At this point, I
- 3:02set the level to the highest, Max, and
- 3:04to be fair, I set Sonnet on the left to
- 3:06Max as well. By the way, I wanted to
- 3:08provide a good prompt so that our final
- 3:09product, the output, would be good. I
- 3:11didn't want to test the models in just
- 3:12one direction. That's why we tested them
- 3:14with different scenarios in three
- 3:15different directions. In both, I asked
- 3:17for a 3D Mars scene that runs in the
- 3:19browser. You drive a rover, a dust storm
- 3:21starts, day turns into night, and so on
- 3:24.
- 3:25In the second one, I wanted a simulator
- 3:27showing the world 1000 years later.
- 3:30We wanted a 3D globe and a time bar to
- 3:32be included.
- 3:33I wanted it to show the changes in the
- 3:36world as you slide from 2026 to 3026.
- 3:40Our third direction was to have both of
- 3:42them create an AI interface that acts as
- 3:44if it were conscious.
- 3:46You type something, and the robot's
- 3:48thoughts and mood change live on the
- 3:50screen.
- 3:51I will check if the work is finished
- 3:53using the same list again. We will look
- 3:55together at whether it opens in the
- 3:57browser, if the interaction works, and
- 3:59if the visuals are smooth and fluent. I
- 4:01put it into the prompt screen. You can
- 4:02try the same thing on your own computer
- 4:04. We started both at the same time. I
- 4:06sped up the waiting parts, but I left
- 4:08the important moments for you.
- 4:10I would like to state one thing from the
- 4:11beginning. I turned off command
- 4:13execution permissions on both sides.
- 4:15Meaning both models could only write
- 4:16files to their own folders.
- 4:18They didn't touch anything else on my
- 4:20computer. And this made the test fair.
- 4:22Because both had to work with the same
- 4:24constraints. Now, I started both at the
- 4:27same time, and the first difference
- 4:28appeared in speed. The Flash model
- 4:30finished all three tasks between 8 and
- 4:3213 minutes. Sonnet 3.5, on the other
- 4:34hand, thought for much longer.
- 4:36It delivered the files between 38 and 48
- 4:38minutes.
- 4:40There is also something interesting here
- 4:41. Since neither model could execute
- 4:43commands, I couldn't test the code in
- 4:44the browser.
- 4:45The Flash model wrote this clearly. It
- 4:47said that since it couldn't run and
- 4:49verify the file, it checked the code
- 4:50manually and finished the job that way.
- 4:53Sonnet, however, called an assistant
- 4:55agent to check its own code, and that
- 4:57agent kept asking for permissions.
- 5:00I didn't grant those permissions and
- 5:02stopped it after one hour. So, Sonnet
- 5:04worked much more meticulously, but it
- 5:06was also quite slow. The Flash model
- 5:08approached the tasks more practically.
- 5:10Let's look at our Mars example first.
- 5:12This is the version Sonnet made. It set
- 5:14the scene at Jezero Crater. By the way,
- 5:16I'm hearing this name for the first time
- 5:17too. I had no idea. There is a small
- 5:19panel on the top right, by the way. It
- 5:20shows all the characteristics you can
- 5:22think of on Mars, like local time,
- 5:24temperature, wind, visibility, distance
- 5:26, dust density, and so on.
- 5:28Now, as you drive the rover, dust kicks
- 5:30up behind it. It added a nice effect
- 5:32like that.
- 5:33Uh, there is also a feature where we can
- 5:35turn the headlights on and off, by the
- 5:37way. You can also change the camera
- 5:39angles however you like. It ended up
- 5:41with quite a nice design like this.
- 5:43By the way, I think it did a pretty good
- 5:45job with the lighting and sky color. I
- 5:47mean, it looks realistic. Now let's see
- 5:49how M 3.1 Flash handled the same task.
- 5:53By the way, it named its vehicle RS7 as
- 5:55well.
- 5:56It placed eight measurement points
- 5:57throughout the map.
- 5:58You collect them as you go near them.
- 6:01So, it actually added a little game that
- 6:03I didn't ask for or specifically add to
- 6:05the prompt.
- 6:06There is a dust storm effect here too,
- 6:08and it even shows the storm's intensity
- 6:09and such in the top right. It also has
- 6:12three different camera angles, and it
- 6:14even added wind sound effects.
- 6:19By the way, the graphics are simpler and
- 6:21more angular, but I can say it actually
- 6:23thought about more things when you look
- 6:25closely. It added a few more features.
- 6:28It added its own personal touch.
- 6:30So, when we look at the Mars section in
- 6:32summary, both did a good job.
- 6:34Sonnet has better visuals, but we could
- 6:36say Flash was a bit more generous with
- 6:38the gameplay.
- 6:39I didn't see a clear winner, but I'm
- 6:41curious what you think. Write it in the
- 6:43comments. Let's move on to our second
- 6:45example right away. That is the World
- 6:46Simulator. Now, this is the version from
- 6:48Sonnet. As you can see, the continents
- 6:50are clearly visible. I'm showing you
- 6:52this now because I'll show the other
- 6:53model's version in a bit. You will
- 6:54understand what I mean.
- 6:57It displayed the lights, the bars, the
- 6:59temperature levels, sea levels, glaciers
- 7:01, and desert areas much more clearly.
- 7:05We can even see how Africa slowly
- 7:07yellows as time progresses.
- 7:09It also specified this in the top left
- 7:11corner. It wrote that this is a
- 7:13fictional scenario, not a prediction.
- 7:15I didn't ask for that, but it added it
- 7:17itself.
- 7:18You know that Claude's models are a bit
- 7:20more serious about these ethical issues
- 7:22.
- 7:23Now let's look at Flash's globe. The
- 7:25panel on the left is nice. It actually
- 7:26did a really good job, by the way.
- 7:28It shows the numbers and reflects the
- 7:30settings well, but the main thing we
- 7:32want to see—the Earth itself—is not
- 7:33fully visible.
- 7:35As you can see, it's covered in clouds.
- 7:37So, we can't fully see the continents. I
- 7:39mean, it looks more like Jupiter than
- 7:41Earth. I can't make out the continents.
- 7:43Of course, not wanting to give up until
- 7:45we got what we wanted, we decided to fix
- 7:47it and went back to the drawing board.
- 7:50This time, I didn't just tell Flash to
- 7:52fix it. I asked it to first read the
- 7:54code, find the source of the problem,
- 7:56and write a plan.
- 7:57Without changing the code, I had it
- 7:59execute the plan it wrote itself, and
- 8:00this is the result.
- 8:02Continents and oceans are clear. As time
- 8:04progresses, Africa yellows, the
- 8:05coastlines change, and the city lights
- 8:07turn on.
- 8:08Now I'm moving on to my third example.
- 8:10Our robot interface that has its own
- 8:13consciousness. Let's look at Sonnet's
- 8:15first. It gave the robot its own name.
- 8:18Very cute, by the way.
- 8:19Its face changes as you type something.
- 8:21On the right, you can also see seven
- 8:23emotion bars. Things like joy, curiosity
- 8:25, love, sadness, anger, fear, energy,
- 8:27etc. You see brain waves underneath, and
- 8:30an inner voice below that. So it's like
- 8:32the robot is thinking for itself. At the
- 8:34top of the screen, there's a section
- 8:36that says "thinking" and counts my
- 8:37memory as three messages. They've also
- 8:39added some specific questions at the
- 8:40bottom for the user to ask. So you can
- 8:43actually see how it performs without
- 8:45even typing anything. Now, let's take a
- 8:48quick look at the Flash robot.
- 8:50It's named it "Curiosity 7." It has mood
- 8:52and energy bars.
- 8:53It selects the most important word from
- 8:55the sentence you write.
- 8:56It also displays its intent, topic, and
- 8:59memory on the screen.
- 9:01And you can watch its stream of thought
- 9:03live at the bottom. I mean, it's simpler
- 9:06than Sonnet's, but it does more or less
- 9:08what we want it to do.
- 9:10By the way, not a single call goes
- 9:11outside in either of them, and
- 9:13everything runs within the page. It's
- 9:15already a system we built locally. Now,
- 9:17let's look at our results together.
- 9:20First off, Sonnet 3.5 ran for
- 9:22approximately 38 to 48 minutes.
- 9:24It burned through about 7 million tokens
- 9:27.
- 9:28I'm calculating the cost via the API,
- 9:30because since mine is included in a
- 9:31subscription, it doesn't actually cost
- 9:33me money this way, but the result is
- 9:35still the tokens spent.
- 9:38By the way, these numbers are for all
- 9:39three tasks combined. Now, let me look
- 9:41at the M 3.1 Flash. Let's say a maximum
- 9:43of 15 minutes per task.
- 9:45It spent 17 million tokens in total for
- 9:47all tasks, but the cost is quite cheap.
- 9:50Now, Flash spent more tokens, but since
- 9:52the tokens are cheap, the total for the
- 9:54three tasks is 14 dollars and 44 cents
- 9:56less than Sonnet.
- 9:57By the way, it actually got the world
- 9:59wrong at first, but the amount of tokens
- 10:02we spent to fix it was 9 million, and
- 10:04the cost for that part was separate.
- 10:06Also, while fixing the world, we first
- 10:08had Flash create a plan.
- 10:10And then we had it execute it. So, when
- 10:12you explain it step-by-step like that,
- 10:14it performs better. Of course, we don't
- 10:16need to take this single task as a
- 10:18benchmark, and results might differ for
- 10:19other tasks, but in general, the picture
- 10:21is very clear. So, here is my takeaway.
- 10:23Let's remember the architectural system
- 10:25I described at the beginning of the
- 10:26video. For example, you can use a model
- 10:28like this M 3.1 Flash for daily routine
- 10:30tasks. It's much cheaper than Sonnet. I
- 10:32, for example, use Sonnet in my agent
- 10:34architectures for these daily routine
- 10:36tasks.
- 10:37But here, for instance, the Flash model
- 10:38is quite cheap and works more or less
- 10:40the same way as Sonnet.
- 10:41Even though it spends more tokens, it's
- 10:43several times cheaper. This can
- 10:44definitely be added to your agent
- 10:47architecture. If you have a more complex
- 10:49task, then a different model can
- 10:50definitely be used. The whole purpose
- 10:52here is to try and reduce our costs as
- 10:54much as possible.
- 10:56For example, when I think about my own
- 10:58system, I can use expensive models for
- 11:00planning or very difficult tasks, while
- 11:02assigning routine or repetitive tasks
- 11:05that waste tokens to cheaper models.
- 11:08If you want to try the M 3.1 Flash in
- 11:10Cloud Code, you can find the link in the
- 11:11description and the pinned comment.
- 11:14You can also run it on your own computer
- 11:16with the exact same setup I used. And if
- 11:18you wish, you can also use this model in
- 11:20Minimax's own interface. It even has a
- 11:23separate implementation within Minimax
- 11:26Code.
- 11:27But of course, if you are already using
- 11:29Cloud Code and want to switch to a
- 11:30different model, you can just stay in
- 11:32the same interface and change the model
- 11:33.
- 11:34The decision is entirely up to you and
- 11:36your projects. Which models do you use
- 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.