Gemini 4 Hydrogen Is the BEST Flash Model Yet – Full Test!

0:00 And look at that smart little model. Putting that angle there means these cabinets can be printed sand support. If I put insane in the title here, that's like three or four in a row. Another

0:09 day, another model, this time from Google. So Google has released Gemini 4 hydrogen. So this is the newest model in the flash. I know it's called Hydrogen.

0:20 This is basically a replacement for Gemini 3.8 Flash. And just based off of some initial looks at results, this seems like it's actually pretty darn impressive, especially when paired with

0:31 that swarm feature that we tested somewhat the other day with the carbon release. So, I did not have early access to this model. So, everything we're going to see that has been run for

0:40 today's testing has just happened between the time this released and now. We will also just show how much usage aotment this uses up for the entirety of the tests that we run today. Now, in

0:50 looking at the benchmarks right here, I have to be honest, I don't have great familiarity with S. Sweep, but it does seem to be sweeping both Sonnet 55 and GPT subscribe 61 Terra in this

1:05 benchmark. And then it seems competitive with them in a lot of other ones as well. So, keep in mind this is being compared to Sonnet and Terra as opposed to Opus or Soul and then whatever's

1:16 above that, which would be Astra, Fable, and Carbon. So, this seems like a pretty good model. It is very quick in terms of its speed, even running it on the new X high reasoning effort, which is

1:26 something that was not present, I believe, for Carbon. So, this has a higher reasoning effort. And basically, we're just going to quickly go through some of the introductory things shown

1:34 here, prominently showcasing some of its 3D chops, and we will be doing the same benchmarks, but with slightly different changes to them to keep them somewhat fresh. So, let's start out by taking a

1:44 look at the pricing, which is a pretty important factor here. Then we'll get into some fun and light testing. I will say something that gets on my nerves is basically I'm just trying to find one

1:53 concise blog post that lists all of the information I care to go over and there's always like three different announcement posts for the new Gemini models which is kind of as difficult to

2:02 parse as someone looking for information. Now with that some of the interesting things about this model as seen in the deep mind specific announcement post this is adapted from

2:11 Gemini 4 carbon. So, interestingly, this is like something we can expect to have performance like its bigger sibling, which did just come out publicly a couple of days ago and is a rather

2:21 impressive model. Additionally to that, the pricing on this is not discounted till a certain period in time like Carbon is. So, if we scroll down into the pricing specifics for these models,

2:33 you'll see that carbon right here has a specific price that is good until November 30th and then it becomes twice as expensive as well as argon. Now, Hydrogen does not. It's just a $1.50 per

2:44 million in and $7.50 per million out. Now, I do believe Sonnet and Terra are a bit cheaper. Okay, so maybe not. So, Sonnet is $2 per million in and $10 per million out. So, it's actually more

2:57 expensive on all fronts. Okay, so I was wrong there. Huh. So, GPT61 Terra is a dollar per million in and then $5 per million out. So this sits in between the pricing of Terra being the lowest,

3:10 hydrogen being the middle, and then Sonnet 55 being the highest in the pricing tier. And this model shares a lot of its technical information in terms of like what it can intake as

3:18 supported data types with Carbon. So it can basically take in everything. And it also has the same overall context window that Carbon has. So that's cool to see. Just a nice healthy one. Again, I don't

3:29 see a specific figure here for the exact knowledge cutoff, but they just show the latest update as November 2026, though I don't believe that means its knowledge cutoff is November 2026. If I am wrong

3:40 there, um, please correct me. So, with the information out of the way, let's basically just start out by taking a look at our tried andrue tests. Now, I do have one running right here that we

3:49 can see. And I did make sure that all of the previous results from carbon, all of the scratch folder, all of the cache check logs that may have been on the system as well as any cache from like

3:59 Blender and GDAU have been completely deleted. So just to kind of make note that this won't be cheating by seeing any previous results from Carbon. I am running right now a change in the

4:08 Seinfeld benchmark where basically it's actually needing to create this as a 3D model which we're going to then 3D print because I think that'll be pretty cool. But aside from that, we're going to

4:17 start out with our tried andrue browser OS test. Now, this is one where I understand it's probably getting a bit long in the tooth. It's like a main staple of the testing, so I don't

4:26 specifically know how to adapt around or replace it, but I have heard the feedback that like some folks may be getting a bit tired of it. So, I will try to find ways to either make it more

4:36 fun or go through it quicker. With that, we are met with a synth wave looking aesthetic. Now, I did see some mentions that this model seems to have better UI taste than some of the other Gemini

4:47 models. They are very textheavy in the UIs they opt to make. So, we'll see if that changes. Maybe not through this, but throughout some of our other ones. All right, we have a welcome thing here.

4:56 Let's just check the big one. Yep. Good. And we also do have a frame rate counter for the procedural background, which is interesting. I'm not in love with that wallpaper. I definitely am with this

5:08 one. I like this stuff, so it is subjective. We do have notifications there. We have sound and we do have a clock with the correct date and time. We have a nice looking dock right here.

5:18 This definitely seems a bit less cluttered than some of the other Gemini results. Especially if this top bar right here wasn't there. This would be a more clean result than I've seen from

5:27 recent Gemini models. With that, let's just check Neon Vice 3D H. All right. Did he just Oh, okay. I'm not quite sure what he's doing. Looks like he's All right. Weapons. One,

5:44 two. Oh, cool. You can three. Two. All right. Let's just uh It's kind of lagging. And I don't know if it's this or like All right, let's

6:00 Well, I stole the police car. All right, cool. That works. Yeah, the the frame rate is dropping pretty significantly. It's possible it

6:12 may be doing something else that's putting strain on the system. This is at this point an aged gaming laptop. So, I don't think I've seen a browser OS GTA result where we can go in the building.

6:24 So, that's good. It's basic, but it's acceptable. There are no different spaces to drift or jump. Oh, look. We can hide in between the building. I don't think that was

6:36 done on purpose, but it's kind of cool. All right. Space is to drift. Yeah. Oh. All right. Overall, acceptable for a flash level model. Gravitas 3D. This is our second game

6:50 where it has to make something creative. Okay. It kind of reminds me of the thing that Carbon did. Uh what the tether? Yeah, definitely. Uh kind of

7:08 like Carbon's result. These models love synth wave. 180 flip. V is for view. Okay. Red click is to tether. All right. Well, I've had enough of

7:23 this, but it it is definitely unique. Eh, it's still kind of like space vibes though, which they all like as the second game for some reason. That's cool.

7:36 Ether OS. That's actually better than we normally see in terms of like the Neoetch result. Not bad. Better than the Fable 52 file browser. The browser OS. Hello.js.

7:48 Runjs. That's cool. So, it has a little JavaScript sandbox and it can run it. I've been seeing more of that now. pictures. Oh, sketchpad. Oh, we did see code pad. That's what opened our

8:00 JavaScript sandbox thing. So, all right. This is not as nice as they normally are, but nice. Next up, we have our wallpaper lab, Neuro web. All right. I don't like

8:13 those. Chromatic plasma, digital rain, and then auroraflow. Cyber terrain is definitely my favorite here. Then, finally, our special feature. Oh, wow. So, it's automatically

8:28 taking checkpoints and it just brought us back to one. That's actually interesting. Sometimes we see that tile all windows. Yeah, definitely acceptable. And we'll

8:38 wait to see what more we can get. Oh, that's actually kind of cool. All right. I'm glad I found that. So, that is more of a demonstration of our special feature that it's just taking snapshots

8:47 throughout the duration of our testing. Very cool. Now, I should mention that the browser OS was not run with the swarm feature. I wanted to at least try one test without that. Now, every other

8:57 subsequent test we're going to see, including the skate game, which we're going to take a look at right now, was run using this agent swarm feature, which at least with Carbon, made a

9:05 massive difference in the overall output. So, with that, let's take a look at our C++ skate game. Okay, it's a little less graphically exciting than I had hoped it would be. However, it does

9:16 seem to have a bit more shape to the skater than we normally get. I I'm gonna say this board and the wheels are actually moving. This is one of the better skateboard models I've

9:29 seen in this test. Overall, it's a bit cartoonish and kind of low fidelity. However, the camera movement is interesting and the playability seems acceptable. It does

9:41 splash when we go on. Oh, all right. Let's replay that. The the skater looks like it's made entirely of hot dogs. So, all right. Can we hit pedestrians?

10:02 A something I don't like seeing is one, we can't hit the pedestrians, but two, the player model and the pedestrian models are different. So, that just breaks the

10:11 immersion, which I really don't like. You know, I like these. I'm kidding. But it does break the immersion. So, okay. I don't know like what this boost feature is hold space for a bigger

10:24 lollipop. Okay, we just won 78 mph there I saw. So maybe an electric skateboard. That's cool. That angle there was decent. We also get some arm movement

10:40 when we're actually like kicking. So, overall, acceptable for a flash model, but not really at the level of polish that I'd hoped to see. Maybe I will give this a

10:53 follow-up here, and we'll come back to it like at the end, but I would have liked to have seen that just a bit nicer. All right, so I gave it a tiny list of some things I want to

11:03 specifically see in the improvement, and then we'll come back to this when it's done sometime later on and see how well it's able to improve upon an initial result. So, next up, we did the robot

11:12 arm test. And this was interesting because I think I made a mistake in running this on extra high mode. I did not put this on the agent swarm mode, but it was on the highest reasoning

11:22 effort. And basically, it took about an hour to get this. It was incredibly slow between moves, and I would attribute that to myself opting to run it on the highest potential reasoning effort. It's

11:33 sometimes hard to figure one to do with the flash models. Like with Fable, I ran this on medium, and I honestly don't remember what I did with Carbon. So, it did a good job. It was very precise. It

11:44 thought through a very deep amount between turns. It did move the car in some wrong ways where basically it would nudge it, and then the car would become a bit more difficult to grab. However,

11:54 it did ultimately pick it up, move it over to the other side, and then it put like the arm in like this victory position where it's just straight up. I notice some of the models do that when

12:02 they're done and some of them just return the arm to the position where they started. So, it's just a weird thing that I don't know with that. I also asked it to make an infographic of

12:13 how exactly it approached this, some of the techniques it used, and that's basically what we see right here. I did ask for a high-end motion infographic just talking about how exactly it went

12:23 about this because it did take an hour. So it was pretty slow, but I would imagine on a lesser reasoning effort, assuming that it did properly succeed in the task, it would have been much

12:32 faster. So it's just interesting to see, but overall it was a success. So next up, we have our Subway FPS test. However, I have taken some suggestion from comment I've seen recently.

12:43 Someone's been saying I should add in a basically rocket launcher in here. And additionally to that, I said the RPG should affect and damage the environment and train as well if it is used. So just

12:54 some slight changes which will make this hopefully a little more fresh but still maintain the same thing that we can actually compare previous results to. So with that let's take a look at this and

13:04 I will be interested to see if the environment does get like decently damaged if we use the RPG and we'll get some look at UI here as well. Okay, it's overall probably too simple of a card to

13:15 really definitively state UI. All right, this is good on first glance. I'm impressed with what I'm seeing. Good. Right click does zoom and to a nice degree. And we have Oh, that's the

13:28 rocket launcher. We'll save that. Let's take a look at the zombie models. Not bad. They're not just simple blocks. They're not like hyperdetailed, but they are decent.

13:46 Are there bullet holes? Yeah, there are. Good. And oh, because the environment can be damaged. So that's pieces of the wall coming off. That's cool. All right. Train's here. Train looks

14:01 good. Oh. All right. Whoever had mentioned that in the comments, fantastic suggestion. Thank you. Good. We have swaying handles

14:10 on the train. Okay. It just froze up a bit. That's all right. Now it's unfrozen. Let's reload everything. All right, I'm just going to come out

14:24 with a with a vengeance. Interesting. Oh, kind of like we saw with Carbon. The second round here is significantly

14:45 more aggressive than normally round two is in these games. I'll say that that's this is actually quite fun. Different sounds depending on the material we shoot

15:00 like that. That's just because of the the damage um prompt though. We only said that should happen with the RPG. So, respect. Let's see if it The problem is if I put insane in the

15:18 title here, that's like three or four in a row, which I don't want to do, but I also don't want to not give it to a model that may be deserving of it just because I've used it before. So,

15:31 I'm kind of I don't know what to do. Can we blow ourselves up? No. Good. Good. This is way more fun with the RPG and the damageable

15:47 environment. So, the the individual who suggested that, thank you. Hello. I want to see if I can do damage to this sign. No. All right, that's fine. All

16:08 right, I'll go one more and then we'll Oh, okay. That's Oh, there's even a front of the train. That's something I never really check. But we can move about the entirety of

16:22 the train, so that's that's nice. Oh, good. There's a there was a much bigger one and then we died before we got to see it. All right, this was this was a lot of fun and keep in mind some

16:41 of the crazy things like the lights that we could shoot out and things like that were because I specifically told it the environment needs to be able to be damaged. So, still though that's a fun I

16:52 think addition to that prompt and I'm happy that we can make it a bit more fun but still be able to compare it to previous runs. Nicely done. So, I had it do the Blender and GDO pool party game.

17:04 It made something that actually looked pretty respectable. However, it was kind of cookie cutter. So, I added on sick. Now, add some logic where if I get a lower score than the other players, it

17:13 turns into a mega sick beat him up cartoon style. Outrageous. Added sore loser brawl. An outrageous cartoon 2.5D beat him up showdown that erupts whenever you get outscored by the other

17:23 players. So, we may actually be able to see multiple different styles of graphics here as well, which could be fun. So, on first glance, the graphics here and everything, this is not bad for

17:33 a flash model. So, just do All right. I don't love this UI right here, but okay. The water's well done.

17:51 X diver. Okay. And again, if we get outsc scored, it turns into like a beat him up. Oh, so if we press F, we can snap and brawl. I will do that in a second, but let's just look at some of

18:05 the This is overall not bad. And this was the first try for this. So, it made this in one go. And then it added the fighting logic. All right. Sore loser, bro.

18:17 That's world star. He said, "All right, how do I Okay, that's so weird. Oh, knock. Okay. LFQ.

18:58 And I did ask for like a cartoon style. So that's I'll try one more just to dive normally. Hold on a second. Bounce power highest. Yeah, that'll be

19:26 good. Overall, this this was definitely an acceptable result like graphically and this was using again Blender and GDAU

19:42 and then just adding in the random fight thing was just a fun little like interesting thing. And I like the way it turned it cartoon in 2.5D. Not bad,

19:55 especially for a Flash model. I don't think it's done. But the thing I hate is the textures and like the brick disgusting ugliness of these buildings.

20:10 All right, so I sent a follow-up basically saying the like the aesthetics are ugly. It's better, but like it's not cohesive in design. All right, so our SC redo is done, and this is the third try

20:22 that we gave it. After seeing what it had done the second try and getting angry, I sent it back just saying it seemed like a bunch of assets put together without any cohesion. And then

20:32 it made this. It's definitely more modern and the style is a little more cohesive. Let's take a look. How are the physics? Not good. All right. Well, let's just go

20:48 look at the waterfront maybe. Interesting. And for a flash model, I guess even though this was after a couple of tries, just asking it to make it better, this is definitely more

21:03 respectable. Do we have splash effects? Of course. Good. All right. Overall, respectably done. I want to replay that. I don't know if

21:22 it will show the entirety of that replay. Yeah, it's going to, isn't it? Nice. It didn't show the pedestrian we hit, though. All right, so the Seinfeld

21:39 test has completed. And again, this is to basically create an STL, a 3D print file of this with respect to something that can reasonably be printed on a standard consumer 3D printer and around

21:51 100 millm x 100 millm, which I think is 5 in or so and however high it should be to keep in scale with that 5x5 dimension. So, we notice a couple of things in the repo here. Just basically

22:02 it made all of this trying to identify the floor plan properly and space everything. We see that these bounding boxes have even been drawn around it to just help itself understand a calibrated

22:13 3D isometric projection overlaid onto the illustration. Things of that sort. So, it definitely went at this a bit scientific, which I'm okay with. And this is kind of giving me the same feel

22:23 as what I saw when the robot arm test was happening. We have a verification sheet. Okay, that's actually quite good. renders. And again, it would have

22:38 rendered this in Blender. So that's what we're looking at right here. I'm admittedly a bit concerned just at how some of this may print sand supports, but that looks like an

22:52 exceptionally exceptionally well done transposition of that source image. Better than what we see when we just ask them to make 3JS. So that's interesting.

23:04 Let's print it. Yeah, we'll print it in blue because that's the color that's loaded in that printer. So, I love AI. I know a lot of people hate it. I get it, but like I

23:15 don't. So, look at that smart little model. Putting that angle there means these cabinets can be printed sand support. Even the arches it's put in the doorway and stuff

23:28 like that. So, this definitely did some thought about like even the way the chair is actually tapered up from that cylinder right there. That shows some level of understanding of like, okay, I

23:40 need to make this printable. I'm going to throw these on Etsy for like 20 bucks. No, no, I uh I no longer have an Etsy store. It's a whole thing. But Gemini actually played a part in the

23:52 decisions leading up to that no longer being a case. Don't ask for more info. All right. So, it's probably going to be best to just show this with B-roll because I don't know how well it will

24:00 show with me holding it. But the 3D print here that this had done of the Seinfeld apartment based off that source image is absolutely fantastic. Now, keep in mind I printed this on the worst

24:12 quality setting just to get it done faster. So, this was like a 24 layer height, which is not very fine grain. And even then, some of the stuff came out actually acceptably considering what

24:22 you would think. Overall, this was definitely the most dimensionally accurate replication of this. Now, normally we just have them recreate that image in 3JS. This created this in

24:34 Blender, so keep that in mind as well. It may have been a better tool for the job, but overall, this thing really is it's exceptionally well done. I'm very pleased with this. It printed with

24:45 absolutely no supports to the point where when I actually put this in the printer slicer program and sliced it, which is just preparing it to be printed by the printer, it didn't say that it

24:54 needed supports anywhere. So, that's cool as well. It did a very good job here. And I'm kind of happy with this. I want to do more 3D printing tests, and I think this was a fun one with something

25:04 that we're used to seeing. So, overall, this is pretty cool. If you're interested purchasing this, link in the descript. No, I'm kidding. All right. So, we see something right here that is

25:13 finishing up. And this is a random wildcard test that I wanted to run just to give it something that's never been seen in any training data and also perhaps something that's a bit of a

25:23 pertinent task considering everything that's going on. So, I'm not going to read through the entirety of this prompt verbatim. However, I essentially looked up the current Wikipedia page for an

25:33 ongoing scandal with a very large company. I had the names changed so that there was no specific like onetoone naming here and then I basically told this to take this bit of information

25:44 turn it into a C++ game without using ray in 3D. So this is a difficult 3D game design task using C++ with no dependencies pertaining to a very very current and hot topic about a big

25:57 company and some issues that were happening. So, this will be Let's just run it. All right. Above board, an anthology of silence, institutional caution, and

26:09 public optics. New game settings. Oh, wow. All right. Decent UI here for C++. I'm going to keep everything the same. Sometimes music volume is best turned

26:21 down quite a bit just from previous stuff. Back. All right. And I told it to create voice assets. However, it can So, the voice assets definitely do not

26:39 work. E, click to write an apology letter and it fills it in for you as you click. Okay, clever. Okay, so we're supposed to take some

26:53 photos. And remember, this is all C++. So, if you are used to Steve the PC repairman and then see these graphics and are like, "Oh my god, it's it's for a reason." What am I supposed to do?

27:06 Oh, okay. Proceed to next chapter. The >> the voice whatever happened there was a complete and utter failure

27:43 the file. We'll skip that one. Same car. All right. Interesting. So we have to follow this. Okay. And we can speed up. Oh, okay. So, I'm just using WD

28:02 to keep the car in the green zone. Okay, proceed to next chapter. It seems like the things I'm supposed to be doing, it won't Let's just Let's just run through some. It's less entertaining

28:20 than I thought it was going to be. It's just like made it very documentary style and boring. All right. I mean, this was I don't really know

28:49 what exactly to make of this. It was really boring. It made it like really kind of a boring documentary style game. It was dramatic in terms of the music and the scenes. It named it above board,

29:01 which is something that was suggested just in the prompt that I sent it. It was overall interesting. So I think the takeaway of this is this is just a game that it made that's cinematic using C++

29:12 with no Rayib. So all of these dependencies and things like this as we see in the credits are listed here and this was just something I wanted to try because it's a random test that's not at

29:22 all ever going to be in the training data. And overall it was definitely interesting. I don't have as much to say about it as I had hoped I would. So for the closing thoughts I think overall

29:32 that definitely considering this is a replacement for Gemini 3.8 eight flash. This is an exceptionally well done model. The subway game was definitely the most fun one in recent memory due

29:44 very heavily in part to the suggestion from the comments to put a rocket launcher in and then just making it so some of the environment can actually be damaged as well. That really added a lot

29:52 of enjoyment to this game. This was probably my favorite result here to play with. the skate game. It did a decent job in making it better, but it was still just not at the level of like even

30:04 Sonic 55, which is interesting. That model's definitely properly tuned for like impressive 3D graphics and things of the sort. However, I will say something that also sticks with me in

30:14 the same way that the Subway game does here is going to be this 3D model that this created based off of that photo of Jerry's apartment. I mean, this thing is absolutely fantastic. And even looking

30:25 at the Blender model that it made, I mean, this was even exceptionally well done as a replication of the source image. And keep in mind, we normally have this replicated in 3JS. However, I

30:36 struggled to see how it was just so much better done in Blender. I mean, the analysis that went into this, you would think would have been done when having this made in 3JS as well. So, I just

30:48 don't exactly know why this model so heavily destroyed every other model that we've had. try to replicate this in 3D. However, it was nice that it happened and the print was fantastic. No

30:59 supports, no problems and it's a good little like handheld size for like a Etsy sale. Now, the final thing is of course usage. So, every result that I tested today with the exception of the

31:11 browser OS was using the swarm feature and every result in general was run on extra high which is that new higher reasoning effort that we have here. So overall in terms of usage, my weekly

31:23 usage is at 80%, when we started this, or basically prior to running any of these tests. I took a screenshot of what it was at. It was at 84%. So the entirety of everything we saw today took

31:34 4% of the weekly usage, which I think is really darn good. Now, I'm on the ultra plan, which I believe is $200 a month. However, that seems to be like a decently acceptable amount of usage

31:45 considering that that swarm feature when you do enable it is actually listed as a preview and quota heavy. So, something that is even mentioned to use a lot of usage. Using that a ton here, 4% of our

31:57 weekly usage on the $200 a month plan. So, I would say that's a fairly acceptable like amount. I don't know, that's always subjective, but I was happy with the amount we got done. So

32:08 overall, that is probably going to conclude our first look and test of Gemini 4 Hydrogen, the new flash model. If you have any questions, please feel free to leave them in the comments. And

32:18 thanks for watching.