Gemini 3.8 Flash Is HERE – Testing Google’s BEST Model Yet!

0:00 We're getting spam notifications that we have been destroyed and that the system is offline. So, Google has released Gemini 3.8 Flash and this is a review video that I had started filming

0:11 yesterday. However, the clips that I had may have been deleted because of my own error. Nonetheless, today we're going to be taking a look at the new Gemini 3.8 Flash, which I know is probably going to

0:21 be overshadowed by the brand new release of GPT6 Astra. Let's seriously get into it and just take a quick peek at some of the things that are new about 3.8. 8 flash. And before we get into it, do

0:31 feel free to subscribe so we can get that 75K and then moving on to the 100K plaque. And let's get started. Now, something that's pretty interesting about this is just the rapid pace at

0:41 which these flash models have been coming out. They mentioned prominently right here the momentum of 3.7 Flash, which was from 3 weeks ago, and this is the third flash release in only 6 weeks.

0:52 That is pretty significant in a rapid pace of releases. The important part of that is essentially what we see right here where each flash release has been building on top of the previous one in

1:02 terms of performance, but the pricing remains the same. Now, on that point, pricing for this is currently 50% off until the end of the year, so December 31st, meaning it's $3.75 per million

1:14 output tokens and 75 per million in and double those figures after December 31st where we'll at that point we'll probably be on like Gemini 4.7 flash. So, with that, it's also interesting to note

1:26 because the big state-of-the-art powerful models are more expensive in terms of the output. It's really interesting to see what the cheaper and maybe slightly less capable models can

1:36 get us in terms of comparing between price toerformance ratio. The model takes in text, image, video, audio, and PDF. It outputs text. The context window is a little over a million, and the

1:47 maximum output token limit is 65536. So, something I also quickly want to mention is there are these two specific additional flags you can basically invoke when you're using this in

1:57 anti-gravity. Now, I did try this teamwork preview one for the C++ skate game because I had started filming this yesterday and lost the footage. However, I have not looked at the output of this.

2:07 So, I'll be interested to see this, but it seems like when using either this or this boost deep reasoning invocation, the model will perform better than it would were we not using any of these.

2:18 So, we will also try some results with this because it can hopefully get us the higher potential performance out of this model. And with that, let's just get started by hopping into anti-gravity

2:28 where we can see the model is selected and it is on high with fast enabled, which is just I think by default for a limited time and these flash models are very, very fast. So, to begin, we're

2:38 going to try the browser OS test v2.7. I'm trying to invoke this boost thing right here and it doesn't automatically like fill it in which makes me a little concerned because if I do the teamwork

2:49 preview one it already shows up right there. I don't see boost having that same behavior. I'm going to try it anyway. I don't know if it will actually make a difference. But this is fully up

2:58 todate and everything like that. So I may have to do a bit more troubleshooting. But because the browser OS test isn't necessarily a super difficult one, I think it's okay

3:08 whatever it tries to do. And Gemini will normally give us an implementation plan. I have changed some of these settings in here. So it will automatically just approve the plans it generates. So it

3:17 keeps going without needing any real user action basically. All right. So we finally have the file and it just gives us the specific instruction to open it just by running this in the terminal.

3:28 Okay. All right. Again with the synth wave. Every single flash model that I've tested is absolutely in love with like the synth wave aesthetic. crit. Okay, so

3:44 we're getting spam notifications that we have been destroyed and that the system is offline. Didn't that happen this morning with like a a lot of AI?

3:55 Okay, let's just All right. Can we get in some of these? Okay. Whoa. A I I I think the notification system in this browser is maybe a bit off.

4:14 Actually, the Mac OS notification system is equally as enraging. All right. Uh we can get out of the car. I do not see any pedestrians.

4:29 I'm just going to ignore the notification. Hijack vice taxi. All right. All right. These notifications are like All right. It fixed that in like a few

4:48 seconds basically. So, that's good to see. All right. We already took a peek at our GTA game. Is there a right click? Yeah, there is. And it's very, very in the theme of this. So far, I like what

4:59 I'm seeing. We do have a clock with the correct time and even the second counter in the bottom right, I believe. Okay, cool. So, this is our ability to just cycle through the different procedurally

5:08 animated wallpapers. All right. Oh, that's a bit fast for my liking. Obviously, I'm going to have to stick with this one, but those were not too bad. Is this a Okay, that's like a It's

5:19 a sound bar. Interesting. We saw Vice City. Here's Cyberdoom 3D. Could this perhaps be a non-space rocket game? Oh, interesting lack of sound in this game,

5:42 but it's actually really a refreshing second game. It's not like a space asteroid game, which is Oh, all right. All right. What's H All right, cool. And then we just respawn if

5:55 we lose. We have infinite ammo with the plasma carbine. That's actually a refreshing second game. Hyper mail. I love these icons, too. But again, this is probably subjectively

6:06 influenced because I very much like this aesthetic in general. So, make note of that. Interesting. As with all mail apps, basically it has meta references to specific games on this browser OS.

6:16 It's just very interesting how Infernace, that's an actual vehicle, at least from San Andreas. I used to smoke people in those back in the San Andreas multiplayer games online. That was my

6:27 car of choice. Um, sorry for that. Sent, start trash. Overall, not bad in the style of this browser OS. Now, I had not actually checked the start menu right here. Okay, cool. This looks good.

6:40 Again, everything is very in the theme of this browser OS. Nice. Hey, okay. I don't know where got like That's okay. Oh, cuz Ether holographic AI companion. Oh, no. All

6:57 right, we'll look at that. Quantum code and scratchpad. No, it made a code compile. You know what I mean? It's like a it put a little code editor in here in lie of just a

7:08 notepad. That is awesome. Now, if I knew how to code, this thing would be fun to test. I'm kidding. All right. So, let's just uh let's try something else. Can we make

7:18 a new file? All right. Well, it did kind of do some of this snake game that I just wrote myself by hand. Next up, we have our special feature, which is this Nova AI thing. Was there

7:30 settings or anything where we could have more access to the wallpapers? No. Okay. What was the cascade windows feature? Let's just Oh, okay. Just to clean them up. All right, let's check our special

7:41 feature, which is some form of AI assistant. Oh, that's sick. Okay. See, I like this sort of like synth wave, and I understand it's subjective, but I really like it. So, it comes through in the

7:51 reviews for better or worse, but I think that's a cool animation. Greetings, operator. I am Nova. Okay. Does this thing speak? Cool. All right. I'm happy with this.

8:02 Overall, I like this. It's very synth wave, which this Flash family of models has liked for generations now. This thing's grandfather liked these. Its great-grandfather liked this aesthetic.

8:15 So, all right. Next up, let's just move on to one of the other tests that I had run yesterday. However, I did not look at these. This was run using that workflow invocation flag. Okay, not this

8:26 one where I said it was garbage. This was one that was just We'll look at that, too. But here's one because I wanted to have a comparison of what it would produce when given this teamwork

8:36 preview scale versus just pasting in this self-contained C++ skate prompt. This is the one with the NYC map. Now, I have not at all even looked at this or played it. So, all right. So, this is

8:46 our you whoa. It's got potential, but it's lacking in um like you know, togetherness

9:03 would be the term that comes to mind. Okay, let's Something I want to actually make note of though is I know don't get me wrong, this is not good. the way the pedestrians are actually walking and

9:15 moving. This is arguably one of the more decent pedestrian implementations in self-contained C++ for this game that I've seen. Also, the curvature of the board's not bad. Our rider does have

9:27 movement, although the arrangement of their body is contorted in a way that makes me uncomfortable. Let's see what's up with this. All right, so I'm sending in an image of this and just saying,

9:37 "Hey man, could you tell me what this is?" as it's definitely just not quite right and it should understand, hey, like the environment isn't rendering. All right, so although it kind of

9:45 spoiled it, this actually seems to look somewhat promising. Although it does look like the rider's body is still a bit concerning. Something I want to note is this definitely is prudent to use

9:55 with that team sub agent skill because it definitely produced a better result than it would have otherwise based off of what I had run with the same prompt when not using that skill, which just

10:06 produced something pretty poor. This is actually this is pretty darn good actually. I think now we have to keep in mind this is the flash model but it is

10:18 benchmarking extremely high higher than every other model. So I think it's beating GPT Astra but I don't know that it's beating Muse Spark 1.3. So I mean aside from the body of our player being

10:31 really messed up. The pedestrians look good. The scenery is not bad. There are buildings in the like distance there which come in. Let's see what we do if we go on this ramp.

10:43 That was that was actually kind of that was kind of cool. I didn't want to get like teased for getting over excited on that. All right. This is actually to be honest with you, this is significantly

10:53 better than all previous Flash models have done. I don't know how much of this can be attributed to using it in the sub agent workflow thing, but even the models of the taxi and things like that

11:02 are not bad. It's a bit sparse in terms of the population. There's not a lot of pedestrians walking around and things like that, but the steam coming out of the manhole, even if the manhole is a

11:12 bit sideways, that's okay. This is really well done. This actually kind of slaps like Oh, it even has um like a lean thing. Why can I not speak

11:28 today? There's been so many models dropping. I like the underside of the board. The wheels are defined well. It even has like grip tape material on top of the board. Oh, can we run into

11:38 pedestrians? No. All right, that's okay. Would have been nice, but All right, this is surprisingly nice. So, next up, this is another one that I

11:49 had run. However, I had not done this with any special workflow, and it's I swear you're going to think I'm lying. Even before I even looked at the result for this, I sent it a follow-up saying,

11:59 "This is very low quality. make it better. I swear I have genuinely not looked at a single version of this website. This is the slapest watch website. So, let's just see what it had

12:10 done here. All right. And here is what it would have made for the watch. Oh, okay. Well, I was right. Unfortunately, it just this seems to just elude this model. It's

12:26 Oh, yep. Okay. So, we can start to see some of what's gone wrong because the actual face there now does look fairly decently put together. The big problem was just the arrangement of everything

12:38 here was quite disastrous and the movement their gears are spinning along the wrong axis. Yeah. All right. Auto pulse. That's maybe not

12:51 sound mechanical escapement activated four hertz. Okay. So, you can actually hear the movement of the watch. I've never seen that in my life. However, unfortunately, this is quite a troubled

13:02 result. So, admittedly, it has been a couple of days since I last filmed this video, but we're just going to pick up right where we left off with the subway FPS. Though, I don't know if we left off

13:11 on this. This is the next thing we're going to take a look at. I did run this using this boost feature right here, and it took a little bit of time and then culminated in just giving us this

13:20 command to run right here, which should give us our subway FPS. So, it'll be interesting to see how this does, especially because it used the boost. Okay. So far, I am a little concerned.

13:29 Hey, I'm less concerned now. This is actually not bad. >> All right. It's definitely It's in line with what a flash model. It's in line with what a flash model, I

13:46 assume, would make. It's not the highest fidelity, however, it is decent. Oh, we have multiple weapons, too. All right. All right.

14:01 This one's definitely more difficult. Most models I do this with, like it doesn't get super tough until maybe the third wave. This one, the second one was basically

14:12 kind of it. Hey, cool. It actually modeled the inside of the train. That I'm actually I like that. Uh-oh. I don't know if we can get back on the

14:24 That's actually cool. You can kind of see them like jumping out of the train at least. Oh, okay. Overall, it's not bad, but at the same time, I would like to see better. So, I'm giving it a photo

14:34 of the Fable 5.1 result that I just screen capped and cropped a little so my face wasn't in it. And I'm telling it basically, give me a pixel perfect replication of this. I don't know how

14:43 well it will be able to do this, but I want to see if it responds well to like visual guidance. All right, this worked for over an hour to pixel perfect replicate this based off of the Fable 51

14:54 result. Okay, this actually did a nice job. It really actually it didn't see the enemies from within the screen cap I gave it. However, if we actually just look at the scene as a whole, it did

15:06 mirror a lot of the elements of it. And the fact that it worked for over an hour, I think is actually pretty impressive. Okay, so the train is the same as it was before because the train

15:16 wasn't in the screenshot, but Oh, okay. And then we lost. I just quickly want to actually take a look at this scene just before we get killed. All right, definitely a better result when given a

15:30 high-end reference to replicate. And it worked for a very, very long time. I think this is the longest I've ever had a flash model work on something. So, keep that in mind. Interesting. This

15:41 next test is something that involves controlling a physical robot based off of what it can see from an attached webcam. This is something I first gave to Fable and Astra in my previous

15:50 head-to-head video of those two where it will essentially have some information about the robotic arm it's controlling. It will have a camera placed in a suboptimal location and then a task.

16:01 Basically, use this robotic arm to pick up the orange car and then move it to the other side of the platform. This is something that should test Gemini's pretty good vision capabilities as well

16:11 as just dealing with suboptimal environments for testing a physical robot. So, let's try this. All right, so in 36 minutes, this successfully completed the task. And interestingly

16:22 enough, this is quicker than GPT6 Astra on Max. That took around 40 minutes. And then Fable 5.1 on Max, I stopped at an hour and 20 minutes because it was really going very scientific with it,

16:34 but it had still not managed to do it. Now, something that may have almost been a lucky move was this accidentally flipped the car over and then hit it and flipped it back. However, it landed in a

16:44 much more, I suppose, kind position for this to actually be able to move it. It really struggled where it was not understanding the depth perception for quite a while where it kept trying to go

16:54 next to the car and pick it up. Ultimately, it figured out somehow that okay, I need to go over a little more and then it actually wrote a script to just pick and place as it called the

17:03 script. So overall, I was actually quite quite pleased with this result. The speed of the model and some of its vision capability definitely played a part in this becoming a successful and

17:13 quick result. So overall, I'm actually quite pleased with this. We can see a bit more about what it did because it saved all of its work into this walkthrough right here. A bigger TLDDR,

17:22 the summary of completed work. It had a hardware driver to move the arm, a vision pipeline just to take pictures and then use them for feedback, and then this mission coordinator script for pick

17:32 and place because it actually did the entire move in a single script as we saw right here, which was an interesting approach as Fable, well, I won't say Fable cuz it didn't successfully

17:42 complete. It was stopped prematurely. GPT6 Astra just kind of moved a little, checked itself, moved a little, checked itself. This one kind of did it more, as we see right here, in a script. So, that

17:53 was actually pretty interesting. Now, some things we see right here are pretty helpful where it has this initial position right here. This was after it had mistakenly hit the car, flipped it

18:02 over a little, and then moved it. But we can actually see some of the work that went down. Car not detected. Grasp claw 513. And these are all things that it is actually making itself. So, it's

18:12 segmenting these and then bounding them. Car not detected. Transfer to opposite side. Okay. Then, it moved it as we can see right here. And then it dropped it down over there. So, it may have gotten

18:21 a bit lucky with the fact that it kind of accidentally moved the car into a position that was easier to complete this task with, but we have some phase breakdowns here and things like this.

18:31 And then the source code artifact. So, overall, I would say that was a pretty decent and successful run considering that this is a very suboptimal camera view to be able to do this. I mean, if

18:40 we just look at this right here, it's pretty hard to judge the depth perception with this one camera angle right here. So this is something that's difficult and I purposely don't give

18:49 them a good camera angle because I want to see how they do when it is harder in this case. So pretty good. For the final test, I'd like to give them something that requires the use of Blender and

18:59 GDAU to make a fun game. Now in lie of the traditional street eat because I haven't fully done a deep clean of this computer. So I don't want it to find lingering artifacts and be able to jump

19:09 on those to base its game. I'm giving it a task and game for this which has never been done using Blender in GDAU, at least in our testing. This is a flight combat simulator game, something I used

19:19 to run pretty heavily when testing these models maybe a year or so ago. So, I figure it'll be interesting to see how this does when using Blender and GDAU to create this game as opposed to the

19:29 traditional 3JS result. It is again on boost right here, so it will work for hopefully a little bit of time and give us something fun. All right, we've received our Blender and GDO flight

19:38 simulator result. Okay, let's see. Uh, let's look at our models first. Okay, so far this is a bit simpler than perhaps I would have expected or wanted. It's uh

19:58 Oh, okay. Well, the boxes around the other planes, something that I saw, there was this old like flight simulator game, I think XP or something. Unfortunately, the sound is rather

20:09 painful to hear. So, maybe we'll Okay, shift is like a Oh, that was interesting. It actually included some like uh pull up like that sort of stuff. Overall, I'm I'm not thrilled with this.

20:24 Okay, cool. That at least worked. I'd say it did better with the robot test, maybe. All right, we were getting hit there and we actually did take some damage. The

20:36 thing with these games is they're generally just incredibly difficult to play. It's simpler than I would have hoped to have. And I did run this one on boost.

20:45 It had hung up for a while. So, I had to send it a follow-up and it said, "I apologize." Like it uh one of the headless scripts hung up. So, I don't actually know how long this took.

20:54 However, okay, so apparently 18 minutes. And I did run this with boost. So, make of that what you will. It's fast, but overall I was not as enthused with this result as I would have hoped to have

21:04 been. So, that's going to conclude our first look and test of Gemini 3.8 Flash. Overall, it does seem to be an improvement over 3.7. And at this pace, I would imagine we'll get like a 3.9

21:15 flash pretty soon. So, that would be cool as well. Unfortunately, I can't really do the traditional results overview because I don't have a lot of the results that were run when I first

21:23 started filming this video quite a few days ago. There was just so much activity that happened in between that time and now. But I did want to get this out and I wanted to at least do

21:32 something a bit novel like basically the robotic arm test where it was actually a pretty cool thing and just a more realworld use case. Overall, the C++ skate game test when given the chance to

21:42 improve itself really showed quite a decent improvement and I was happy with that. The model loves Synth Wave. It is very fast and again compared to big expensive Frontier models, it's fairly

21:53 cheap, especially till that December 31st date. So overall, I wanted to just get a video out testing this. I always do like Gemini models a lot. I still use 3.1 Pro a lot for non-coding work use

22:04 cases and things of that sort. So, that's going to conclude our first look and test of Gemini 38 Flash. If you have any questions, please feel free to leave them in the comments. And thanks for

22:13 watching.