Modern Creator
David Ondrej · YouTube

Gemini 4 Neon vs Fable 5.2 vs GPT 6.2 Astra

Google's middle-tier model tops coding benchmarks at a fraction of Fable's price. A 50-minute head-to-head on images, websites, 3D, games, thumbnails and launch videos, plus which plan and CLI to run it on.

Posted
yesterday
Duration
Format
Review
hype
Views
23.2K
353 likes
Part of the collectionThe Claude Fable 5.2 PlaybookEvery Fable 5, 5.1 and 5.2 breakdown, synthesized into one page.
Read the playbook
Part of the collectionThe GPT-6 Astra PlaybookEvery GPT-6 Astra breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

Gemini 4 Neon is a mid-tier model that performs like a frontier one, so it belongs in the daily-driver slot next to Opus 5.5, run on high effort instead of extra high.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You burn through Fable 5.2 limits on Claude Max plans and want a cheaper model that holds up on real coding work.
  • You run coding agents daily and need to decide whether a Google subscription is worth adding to your stack.
  • You have written off Gemini after a year of weak releases and want evidence on whether that still holds.
  • You run nightly security or code-review loops and want a cheap second-opinion model from a different lab.
  • You generate images, video or thumbnails through an agent and want to know which model writes the best prompts.
SKIP IF…
  • You want controlled, repeatable evals. Every head-to-head here is one prompt per task, judged by eye.
  • You need a neutral take on tooling. The video is sponsored by Higgsfield and leans on the creator's own Cloudroom app throughout.
TL;DR

The full version, fast.

Gemini 4 Neon was positioned as Google's mid-tier model, yet it ranks first on CursorBench, first on the Artificial Analysis coding agent index through Antigravity CLI, and near the top of roughly 20 aggregate benchmarks, at 2 to 8 times less cost than Fable 5.2. Side-by-side tests show it winning most image-prompt and arcade-game rounds, tying Fable on web design, and losing to Astra on thumbnails. The practical setup is Antigravity CLI on the $20 Google AI Pro plan at high effort, since extra high adds one or two benchmark points for 60 to 80 percent more cost. Treat Neon and Opus 5.5 as daily drivers and save Fable for jobs that need the extra power.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 01:59

01 · The claim: a mid-tier model at the top

Neon was meant to sit in the middle of the Gemini 4 lineup but is beating Carbon, Fable 5.2 and Astra on several top benchmarks. Roadmap for the video.

01:59 – 06:33

02 · Benchmarks and early demos

First on CursorBench 4.1, competitive across about 20 aggregate benchmarks, first on the Artificial Analysis coding agent index via Antigravity CLI, best on Bug Hunt Bench. Demos: Blender MCP 3D scenes and a tower defense game built in 20 minutes on 2% of a five-hour limit.

06:33 – 09:40

03 · Test 1: writing prompts for image models

Each model turns a short concept into a detailed prompt for the same image model. Neon wins the greenhouse, sci-fi poster and samurai rounds; Astra wins the coffee shop logo and underwater city.

09:40 – 11:18

04 · Sponsor: Higgsfield MCP

Pitch for Higgsfield's MCP and CLI as one place for an agent to call image, video and audio models. About 0.6 credits per image in the test.

11:18 – 13:16

05 · Why Google is winning

Compute, talent and data favor Google. A March 2026 Elon Musk post predicted Google would win the AI race in the West. Includes a tangent on data-center regulation and space-based compute.

13:16 – 17:46

06 · Release cadence and the review skill

Argon, Carbon, Hydrogen and Neon shipped with gaps of 28, 15 and 6 days. Neon now powers a /gemini-review skill used as a different-lab second opinion. Live demo of Neon rewiring that skill, plus a pure-JavaScript motion design showcase. The whole team, including non-developers, is switching some work to Neon.

17:46 – 23:21

07 · Test 2: one-shot coding showdown

Same prompt, all models on high. Web design: Fable and Neon tie, Astra ships a page that doesn't scroll. 3D Tokyo alley: Fable wins, Neon second, Astra broken. Code art: Fable has the best style, Neon the most effort. Arcade game: Neon's orbit game edges out Fable's.

23:21 – 26:37

08 · Where to run it: CLI and plan

Use Antigravity CLI, not the stale Gemini CLI. Google AI Pro at $20 a month includes Neon; Carbon is Ultra-only but Neon beats it anyway. Subscription multipliers: Claude Max on Opus 5.5 is best, Google AI Ultra on Neon second at 25.8x, ChatGPT Pro worst.

26:37 – 27:43

09 · Reasoning effort: use high

Extra high buys one to two benchmark points for 60 to 80 percent more cost, and scores worse on long PDFs. Default to high; reserve extra high for the hardest coding tasks.

27:43 – 33:05

10 · Neon plus Higgsfield: a full build

Neon installs the Higgsfield MCP from a link, then builds a 3D cyberpunk arcade page with a generated hero, 3D prop and soundtrack. The Pro plan hits its limit mid-run; the creator upgrades to Ultra and runs a medium-effort copy in Cursor CLI.

33:05 – 35:49

11 · Writing, honesty and the medium vs high result

Early users call Neon the least sloppy writing model. It ranks first on several office and finance evals and rarely invents answers. The high-effort arcade build clearly beats the medium one. Google 3D demos: a Turing machine rebuilt in Blender and a skiing simulator.

35:49 – 37:18

12 · Training data explains the gaps

Fable leads on protein science because of its training mix; Gemini leads on video, multimodal and long-document work. Neon is the cheapest leading model on a finance benchmark, with Fable at 14x the cost.

37:18 – 38:57

13 · Test 3: YouTube thumbnails

Same image model, three prompt writers. Astra wins two of three rounds with the most creative concepts; Fable wins the third by getting the logo and face right. Fable and Neon prompts look surprisingly alike.

38:57 – 41:25

14 · Cybersecurity loops

Neon is strong on web exploit and bug-finding benchmarks. Recommended pattern: nightly loops that rotate frontier models across every repo, prove each finding, fix it, and have a second model review before pushing.

41:25 – 47:00

15 · Antigravity in the terminal and launch-video test

Setup walkthrough for agy: themes, data-sharing opt out, Shift+Tab modes, /model for effort. The harness forces a plan mode and runs slower than Claude Code. Neon makes a 15-second history-of-AI video, then wins a three-way launch-video test on originality.

47:00 – 50:00

16 · The catches and the verdict

Neon is the best-value workhorse; Astra is odd but strong at creative direction; Opus still wins on short tasks. No full safety report yet. Use Neon or Opus 5.5 as the daily driver and Fable for extra power. Closing Higgsfield Black Friday pitch.

Atomic Insights

Lines worth screenshotting.

  • Gemini 4 Neon, Google's mid-tier model, outscores Google's own larger Gemini 4 Carbon on quality benchmarks.
  • Neon is roughly 3x cheaper than Opus and 8x cheaper than Fable per task on the Artificial Analysis index at high effort.
  • Google shipped four Gemini 4 models in about seven weeks, with the gap between releases shrinking from 28 days to 15 to 6.
  • Going from high to extra high on Neon buys about one Artificial Analysis point for 72% more cost and 53% more time.
  • On long-PDF work, Neon at extra high scores worse than at high while costing more.
  • A $100 Google AI Ultra plan returns about 25.8x its price in API-equivalent Neon usage, second only to Claude Max on Opus 5.5.
  • Fable 5.2 gets a separate, smaller usage pool on Claude Max, which is why heavy users stack multiple $200 plans.
  • When Neon doesn't know an answer, it invents one only 14% of the time, versus about half for Opus and about 60% for Fable and Astra.
  • A model's strengths track its training data: Fable leads on protein work, Gemini leads on video, multimodal and long documents.
  • GPT 6.2 Astra built a landing page with no scroll and a buggy 3D scene, then produced the best thumbnails of the three models.
  • Models from different labs catch different bugs, so a review step from a different family finds what your main model misses.
  • A nightly security loop that rotates frontier models across repos and only reports proven bugs costs little when the model is Neon.
  • Medium versus high reasoning on the same Neon build produced a visibly more complete game, longer music loop and better controls.
  • Opus 5.5 still has a faster time to first token, so it stays the pick for 10 to 20 second tasks.
Takeaway

A cheaper model at high effort beats paying for the top tier

WHAT TO LEARN

Model rankings now change week to week, so test new releases on your own tasks, default to the best value model at high effort, and pay for the flagship only when a job proves it needs it.

02Benchmarks and early demos
  • Tier names no longer predict quality: Neon, marketed as mid-tier, outscores Google's own larger Carbon model on CursorBench and bug-finding benchmarks.
  • Compare cost per task, not just scores; Neon runs about 3x cheaper than Opus and 8x cheaper than Fable on the Artificial Analysis index.
03Test 1: writing prompts for image models
  • When the same downstream tool is used, the agent's prompt-writing skill decides the result; Neon won three of five image rounds on prompt quality alone.
07Test 2: one-shot coding showdown
  • Judge one-shot builds on completeness as well as looks; a landing page that doesn't scroll is a failure no matter how polished the first screen is.
  • No single model wins everything: Fable led the 3D scene, Neon the arcade game, and the two tied on web design.
08Where to run it: CLI and plan
  • Pick the harness Google actually maintains, Antigravity CLI, because the older Gemini CLI has stopped getting updates.
  • Start on the $20 Google AI Pro plan and upgrade to Ultra only after you hit its limits in real use.
  • Measure plans by API-equivalent value: Claude Max on Opus 5.5 still gives the most, with Google AI Ultra on Neon close behind at 25.8x.
09Reasoning effort: use high
  • Leave reasoning effort on high; extra high adds one or two points for 60 to 80 percent more cost and can even score lower on long documents.
10Neon plus Higgsfield: a full build
  • Don't drop to medium for full builds; the same Neon prompt on medium produced a visibly thinner game than on high.
  • Agents can set up their own tools from a link, so installing an MCP is often one sentence to the model.
  • Expect subscription limits to bite during long multi-asset builds, even on a cost-efficient frontier model.
11Writing, honesty and the medium vs high result
  • Low hallucination is a real differentiator: Neon fabricates answers far less often than Opus, Fable or Astra when it lacks the facts.
12Training data explains the gaps
  • Expect each lab to be strongest where its training data is deepest, such as Gemini on video and long documents and Fable on biology.
13Test 3: YouTube thumbnails
  • Prompt-writing taste varies by task; Astra lost on code but produced the most creative thumbnail concepts.
14Cybersecurity loops
  • Run nightly security scans on every repo and rotate models from different labs so each catches what the others miss.
  • Require each automated finding to be reproduced and reviewed by a second model before it is patched and pushed.
15Antigravity in the terminal and launch-video test
  • Harness defaults matter: Antigravity CLI forces a plan step and runs slower than Claude Code even with the same model.
16The catches and the verdict
  • Keep a fast model for short tasks; Opus 5.5 still starts answering sooner, which matters more than raw quality for 10 to 20 second jobs.
  • Use Neon or Opus 5.5 as the default and keep Fable 5.2 for jobs that genuinely need the extra power.
Glossary

Terms worth knowing.

Antigravity CLI
Google's terminal coding agent, launched with the shortcut agy. It replaces the older Gemini CLI as Google's main command-line harness for Gemini models.
Reasoning effort
A setting that controls how long a model thinks before answering. Higher effort usually costs more tokens and time; Neon offers low, medium, high and extra high.
CursorBench
A benchmark from the Cursor editor that scores models on ambiguous, multi-file coding tasks taken from real Cursor sessions.
Artificial Analysis
An independent site that benchmarks AI models on intelligence, cost and speed, including a coding agent index that scores model and harness pairs.
MCP (Model Context Protocol)
A standard way to give an AI agent access to outside tools, such as an image generator or Blender, so it can call them directly during a task.
Recursive self-improvement (RSI)
The idea that an AI system helps build its own successor, so each new version arrives faster than the last.
Subscription multiplier
How many dollars of API-priced usage a flat monthly plan gives you. A 25x multiplier on a $100 plan means about $2,500 of equivalent API calls.
Time to first token (TTFT)
How long a model takes to start streaming its answer. Low TTFT matters most for short, quick tasks.
Agent loop
A scheduled agent task that runs repeatedly, for example every night, doing a fixed job like scanning a repo for vulnerabilities and reporting only real findings.
Resources

Things they pointed at.

23:21toolGoogle Antigravity CLI
24:27productGoogle AI Pro ($20/mo) and Google AI Ultra ($100/mo)
01:59linkCursorBench 4.1
03:39linkArtificial Analysis
15:02toolCloudroom ↗
05:33toolBlender MCP
Quotables

Lines you could clip.

16:43
“How are we accepting that this is normal?”
short reaction line over a striking demo→ IG reel cold open↗ Tweet quote
19:13
“I'm starting to see why people hate Astra.”
funny, self-contained reaction to a broken build→ TikTok hook↗ Tweet quote
27:00
“You literally get like 97% performance on high compared to extra high.”
concrete money-saving tip with a number→ newsletter pull-quote↗ Tweet quote
33:46
“When it doesn't know, Neon says I don't know, whereas other frontier models are more likely to just make something up.”
clear claim viewers will argue about→ TikTok hook↗ Tweet quote
49:10
“Use Neon or Opus 5.5 as your primary daily model right now.”
direct recommendation, no setup needed→ newsletter pull-quote↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

So Google might have just dropped their best AI model in the world and in fact there is a strong argument that Gemini for Neon is actually the best AI model. in the world, maybe even better than Fable and GPT 6 .2 Astra. So this is one of the most underrated and looked over releases from the last 12 months.
Gemini 4 Neon is the latest AI model from Google and was supposed to be like a Opus level release, right? So like a medium tier and then you would have the big class, the Fable, the Astra, the Carbon, but this model is just crushing all the benchmarks. Like on a lot of the...
top benchmarks it's literally outperforming GPT 6 .2 Astra as well as Fable 5 .2 at a fraction of the cost and it's a lot faster and it's from Gemini which is perhaps the most surprising part. So in this video we're going to talk about why Google's middle tier model is beating Carbon and Fable 5 .2, which plan, which subscription and CLI to actually use, so how you can start using it today on your own code basis.
I'll also show you a direct comparison of Neon versus Astra versus Fable on images. videos and coding tasks and we'll also go through the hidden things in the model card and my honest thoughts on this model so if you're serious about AI and if you care about the frontier of the most important technology in the world make sure to watch until the end because this is the most unexpected release ever like literally this has never happened in the AI where a lab was so behind for 12 months and then just dropped four models back to back and as Victor Tallin says the creator of Bend Gemini Neon just beat Fable 5 .2 in nearly everything I tried to write is this supposed to be a cheat model what is going on this is the most confusing part so let me get this clear to you guys i don't even know if it's clear to google okay so look at this right you would think okay carbon is the big boy you know slow speed most expensive the fable class the astro class but then if you look carefully why is neon
outperforming on quality over carbon. Like, yes, hydrogen is like the Sonnet class, right? This is like GPT Terra, Claude Sonnet.
They're kind of like a low to medium class. But why is neon better than carbon? Nobody understands.
And it is the most unexpected model of the year. And it's actually incredible, okay? So first of all, it's number one on Cursor badge.
This is the benchmark from Cursor, which again, they have their own models. Cursor acquired by SpaceX. They have Grog models.
Yet Gemini for neon is absolutely dominating. Fable 5 .2 is incredible. Let me get it straight.
It's... amazing model, I love working with it, but Neon just out of nowhere comes in and boom, state of the art, outperforms Fable, outperforms Opus 5 .5, just incredible, like I don't know what else to say. Then we go to this aggregate benchmark where you can see A lot of the top benchmarks, you might be thinking, okay, David, but people are just going to pick one or two benchmarks.
Well, here is like 20 of them. And you can see like optically without going into each one of them. Yes, Fable 5 .2 is still the best on a lot of things.
But Neon is literally as competitive, if not more competitive on a lot of these benchmarks. But like Neon is like between two and seven times cheaper than Fable. I mean, Fable is very expensive.
I can literally show you right now inside of Cloudroom. I have six different $200 plans for... Claude, because Fable runs out so quickly, but Neon, you can get it.
And again, I'll show you which plan is the best for Neon, but it is way more efficient than Fable and very similar in price. So like what is happening? Because of this, a lot of people are speculating like, did Google or DeepMind crack RSI, recursive self -improvement?
Because as I'm going to touch on soon, they keep releasing faster and faster. From Argon to Carbon, it was 28 days. From Carbon to Hydrogen, it was 15 days.
And from Hydrogen to Neon, it's just six days. And some people are already speculating. sodium or some next model coming out in the next few days from Gemini which is just unseen like after 12 months of everyone mocking Gemini and doubting them now they made the biggest comeback ever in the history of AI, and they're shipping faster than both OpenAI and Anthropic combined.
Now let's look at the artificial analysis report on Gemini for Neon. First off, you can see that the cost to do one of these index tasks is much better than Opus or Fable. I mean, Fable is just super expensive.
If you run Neon on high, and I'm going to touch on the reasoning efforts in a second, which ones you should use, you can see that it's like way, like 3x cheaper than Opus and 8 times cheaper than Fable. It's not quite on this parade of frontier. but basically extra high is here.
However, if we go to the coding agents section here, you can see that on the coding agent index, anti -gravity CLI with Gemini for neon extra high is the best in the world. Literally better than Fable, Opus, Muse, GBD 6 .2 Astra. This model is insane, guys.
I know it's hard to believe. You might be thinking, okay, David, how good can it be? You know, Gemini has been sucking for the last 12 months.
Surely this is just optimized for benchmarks. Surely it's not that useful. I'm going to show you real demos.
I mean, I have multiple demos prepared here, comparisons between Fable, Astra, Neon. We're going to look at the model, how it performs on generating images and writing image prompts, how it performs at doing video demos, how it performs at coding tasks, designs, websites, 3D modeling. It's going to shock you.
There's something that incredible happened. Google and DeepMind and hopefully we'll have Logan soon for part two, Logan Kilpatrick. But this model is much, much better.
So no matter what you're using, you need to change your mindset because a lot of people, humans, we like to be comfortable. We like to know, okay, just tell me what's the best model. This is like one of the most common requests I get.
You know, David, tell me what's the best model. What should I be using? The thing is like this can change from week to week, right?
You need to be willing to adapt. You need to be willing to try these models for your own use cases and make up your own mind. On the bug hunt bench, you can see that from the top.
Top seven models, four of them are Gemini models and Neon is the best in the world. Again, even better than the Carbon, which is supposed to be a bigger model. I have no idea.
Here's some examples of what people build with it, right? So here's Google's own example of the release. And basically this is a 3D, right?
So they gave it a Blender MCP. This is not a video model. This is Neon writing code into Blender.
So like through code controlling Blender through the MCP and building... These 3D designs. This is 3D renders and then it renders the whole scene.
And all of this is done by Neon. The designs, the triangles, the 3D models. Visualization of the four models so far in the Gemini 4 series.
Obviously representing different elements on the periodic table. But this is just crazy. It's really amazing how fast we adapt.
I mean today the Riemann hypothesis was supposedly solved. That's a topic for a whole separate video. But...
It's crazy how fast people see like this and like, you know, don't think it's that impressive, even though this would be completely unheard of just six months ago. Here's what some early developers said. This guy built a tower defense game in about 20 minutes and just used 2 % of his five -hour limit, which is crazy.
I mean, look at this.
Games like this take literally months, if not years, to develop if you're an indie developer. Now it's just 20 minutes with some AI agent. Let me show you my own testing comparing Gemini for Neon.
directly head -to -head against Fable 5 .2 and GPT 6 .2 Astra. So first, let's look at Fable, which I guess it's probably the favorite, right? Anthropic just got IPO.
It's the leading frontier AI lab. So first image, right? And by the way, the way this works is you give these agents a concept, right?
Then the agents need to write a detail prompt and that is sent to an image model. You can see there's different models, right? Flux -free image, ideogram 6 .2, Seedream 5 .0, ReCraft V4 .1, Grok image 2 .0.
So the image models are the same in each. scenario what we're judging on is the creativity of the prompt how good these agents are how good these models are at controlling the tools right so first prompt fable this is the abandoned greenhouse you can see that's the concept and then again fable needed to write the prompt so it's not bad the issue is like the reflection doesn't really match the tree and also it doesn't make sense that there is a tree like in the middle that grew like this i don't know i'm not the biggest fan of it let's look at astra astra i think the image has like more striking the reflection makes sense interesting orange there i would put it a bit better not much better just a bit better and then we have neon which i think this one is probably the best out of the three it's the most beautiful you know the reflection actually reflects the building.
It's a nice floor, bench, good lighting. I would say this one goes to Neon, number two, Astra, number three, Fable. Let me switch to a better view here.
So again, all of them are using the same image model, in this case, flux -free image. Next one, we have the indie sci -fi movie poster. So that's the concept, right?
So here, Fable... nor Astra make any sense. Like, this doesn't strike me as sci -fi.
The only sci -fi is Neon. I think I'm being pretty accurate here, but Neon is the only one that gives me sci -fi vibes. The other ones, I mean, Astra, definitely not.
Fable, maybe, I guess. I don't know. Neon wins this.
Next one, this one is more interesting. We have Samurai in Snow. I like the Fable.
The issue is like the dagger doesn't make much sense. It's not really the katana. Also, if you look at it, the samurai is like very close to this rose bush and he's like up to waist in a snow.
That's kind of off. It's an important detail that's off. Astra.
It's pretty weird. I mean, the focus is not on the samurai, which is a mistake, I would say. He's also bowing down to this bird, but the bird is like walking to him.
So the bird is a lot closer to the camera. So it makes no sense that he's bowing down next to it like this. Neon is probably like the least realistic.
It feels like a video game, but it's also like has the best vibes, I would say. It's the most aura. So again, I probably have to give this one to Neon again with Fable 2nd and Astra 3rd.
Let's look at this one. cozy coffee shop logo now here Astra is the best for sure I mean look at this cute cat in a mug if we judge it on the logo Astra has the best concept by far Neon is pretty bad to be fair here I would put it like tight Fable 5 .2 and Neon here Astra definitely takes the lead here and last one on this example we have the underwater ancient city again all of them using the same image model here Astra is probably the best it's the most realistic I don't like the other two personally I would probably put Fable slightly above Neon but it's close so I don't know overall you can see Neon is very competitive to Astra and to Fable when it comes to controlling image models.
We got the most out of these five. Obviously, the sample size is kind of small. And by the way, this is possible thanks to the Higgs field skills and the Higgs field MCP, which is very OP.
You can just give it to any agent. It's so easy to control. Fully supported with ChairGBD, Cloud, Cloud Code, the Telegram agent, as well as CLI.
Basically, any agent that can use the terminal can generate images, videos, advertisements, creatives, graphics, you name it. Like literally, you can use this for countless different purposes any business or any individual should be using some AI enhanced images or videos and instead of you having to like think of the prompt and you know do all this stuff Hicksfield makes it so easy by giving you an MCP and again you can just tell your agent in plain English hey give me a logo for a coffee shop like here you know or I'm doing a movie about a samurai boom right here.
The agent will do the prompt, right? You literally give it the concept, the agent will write the detail prompt. The agents know how to control these image models better than me and you.
And you can just, you know, run variations with different models. Super easy. And this cost me only like a couple of credits.
I didn't even burn that much. Yeah, I think these three generations cost like two credits total. So 0 .6 each.
So you can literally generate hundreds, probably thousands of these images inside of Higgs field from basically all of the image models and video models that exist. Videos obviously cost more credits because they're a lot more. compute intensive, but it's all in one platform.
And with the MCP, which is my favorite way to use Hicksfield, it's never been easier to create high quality graphics, animations, videos, AI images, just by talking to any agent you prefer. Also, Hicksfield was kind enough to sponsor this video. So if you're generating anything with AI, make sure to go to the first link below the video, click that link and get one of the Hicksfield plans and start generating high quality graphics, images, videos through Hicksfield and Hicksfield MCP.
Again, it's the first link below the video. Now, why is Google winning? This is something...
that is kind of surprising to people and actually Maybe we can find this. I'll see if Opus 5 .5 Max can find this from one of my older videos, like 2023, 2024, where I talked about the first principles, which labs have the most compute, talent, and data.
And I think I also predicted this. If so, it's going to be shown on screen. But regardless, I'm not the only one who predicted this.
In March of this year, Elon Musk said that Google will win the AI race in the West, China on Earth, and SpaceX in space. Now, again, keep in mind that at the time of this tweet, Google was trolling, right? Their best...
was like gemini 3 .1 pro and they were way behind anthropic way behind and falling behind more and more right like really a month ago they were super behind until they started releasing gemini 4 family so this prediction is turning out to be true obviously we'll see what happens with anthropic opening eye is falling more and more behind by the day it's really crazy china on earth has not yet materialized spacex in space again still not yet materialized but actually that one i think it's more likely than china on earth because if you look at the results of the mid turns and, you know, all the leftists, they want to shut down the data centers.
The leftists, they hate progress. They hate technology. All they love is just communism, socialism, and endless poverty.
They're protecting abandoned parking lots, preventing data centers, which generate revenue and boost the entire community in terms of wealth, in terms of jobs. People think it's better to have a dead, abandoned parking lot. If there is so much regulatory pressure on data centers, the only place where it'll be easy will be the space.
And if that's the case, SpaceX will 5X in terms of market. and it will win the ARAs in the real space. So again, that one, I think it's...
Pretty likely over the next two to four years. Google is happening in front of our eyes. It's winning the ARAs in the West.
Obviously, Anthropic is in the lead probably by a bit. But Google has caught up. And if they have some form of RSI, might even get ahead.
China, we'll see. They haven't been shipping much stuff as of late. But it's possible.
And they have the most energy in terms of electricity generation. So it's definitely possible. We'll see if Elon is right.
Let's talk about the release cadence because this is crazy. I already touched on this. Google seems to be releasing more and more.
We'll see if it stops here. Because the worst outcome would be that, again, next 12... months nothing happens I remember vividly 12 months ago when Gemini 3 came out it was the frontier it was the best model in the world everybody was freaking out then you know month later Gemini 3 .1 came out it was a bit better not a huge breakthrough but like ever since then ever since Gemini 3 .1 nothing happened until we saw these four models drop right Argon was fine I even was judging the Argon hackathon in San Francisco but it wasn't like frontier then we got carbon carbon was major deal okay I started using it in my skill but we My skills repo is completely free and open source I'm gonna also link it below the video and you can get all of the skills I'm using and the one I'm using the most right now is Gemini review Okay, so slash Gemini review.
This uses a Gemini frontier model. So currently it's Neon, right? But when I first created this, it was Carbon.
Carbon was the first frontier model and it was great because it reasoned for like 20 minutes and it always found stuff that either Claude or GBT missed. And that's because, you know, in my previous video, I explained how the Gemini family of models is very different than Claude and GBT, which again, the entire Anthropic team is people, descendants from OpenAI.
They worked at OpenAI, they left, they've all followed the same practices. the same training methods, Gemini and DeepMind, they're very different, right? So this is what makes Gemini the perfect reviewer.
But now this skill is superpowers. It basically is enhanced because we got Gemini for Neon, which is way cheaper to run than Gemini for Carbon. It's also way faster and it's better.
Like it's literally cheaper, faster and better. I really don't know what to say. It's on extra high.
I'm running it through the cursor. CLI, but I'm going to change this to anti -gravity. So actually, let me show you how easy that is, right?
So I'll just go into Cloudroom. Let me show you a real demo of Neon. We can do it on high.
It's not an easy task. So find on my computer where this skill is defined. It should be a global skill.
Change it so that it uses the anti -gravity CLI. First... and only the Cursor CLI as a backup.
Boom. So Neon, it actually is similar speed as Opus. Look at this, right?
And I didn't even launch in fast mode. So it's not as slow as, you know, Fable and Astra and Carbon. It's very easy to work with.
And I've actually told everyone on my team to give it an honest shot. And everyone is liking it. Even people who have nothing to do with software development, right?
People who are in charge of hiring, accounting, legal, like they're doing it. They're using Neon and they're liking it more than Opus 5 .5, which was the model that we were using the most previously. So this...
isn't just another release on the side this is literally a potential for your main ai model right i'm not just saying this guys it's very rare where i would recommend all of you to use a main ai model right fable 5 .2 is frontier but it's too expensive that's why you know opus 5 .5 is still the best deal and there's separate fable limit which i hit the soonest which is why i have six cloud subscriptions but neon is actually a great potential for your daily driver like neon is a strong competitor to opus 5 .5 i'm still using both but i'm using neon more and more as i'm learning to trust it more and learning to work with it.
So if you've never tried this model, give it a shot. Like you will be happily surprised. This is the first Gemini model in the last 12 months that is actually usable as your daily driver.
It's really crazy. And by the way, before I show you my own results from the live test on coding tasks, this is an incredible showcase of what this model can do. So this is Gemini for Neon making a motion design to introduce itself.
Pure JS code, pure JavaScript, not a single external asset. Look at this. Okay.
Incredible.
This is insane.
How are we accepting that this is normal? This is an AI model literally creating. If you wanted to recreate this, if you don't have the skills, this is months of work.
If you have the skills to recreate it, this is at least like a week or two of work. And the model can do this like in a few prompts, you know, no problem. It's really insane.
And the amount of people that like understand what's happening, like we're entering the singularity. The Riemann hypothesis has been just solved. The hardest problem in mathematics stood for 200 years.
It's been solved. And people are like acting like everything's usual. If you're not going all in into AI, you're making a grave mistake.
This is literally the greatest technology of all time. And we're seeing acceleration like never before. And Gemini for Neon is...
the latest embodiment of that, okay? So I'm going to show you this, actually, I have this local host. So this is, again, same models, just compared three different models, right?
Fable 5 .2, GPT 6 .2 Astra, and Gemini 4 Neon, compared on the same tasks. So first is web design, best possible landing page for a luxury space tourism company. Let's look at this.
Fable 5 .2, okay? Not bad, nice animation, good color scheme, little bit empty somewhere here. These could be a more high effort, feel a bit lazy.
This is a bit lazy, but I like the... Landing page, you can see the kilometers and the altitude. Long landing page.
This is nice. We got the animation glow. Really good footer.
I like it, okay? Not bad, not the best. It's also on high.
So these are not the max reasoning efforts. All of them are on high. So it's as fair as possible.
Here we have Astra. So first of all, no scrolling. That's a big, big negative.
Like there's nothing else to this landing page. That's it. I cannot scroll down.
This opens the same model as this. Like no real animations or nothing. Yeah, I mean, that's it.
Very lazy. Not good. Let's look at Neon.
What do we have? So Neon has the scrolling, so it's a proper landing page. We got some small animation, a bit less than Fable, I would say, but definitely much better than Astra.
We also got a configurable plane or starship as similar to Fable. We got the animation of the map, which is nice. Fable didn't give us that.
And a nice footer as well. This model is much better than from Astra. So yeah, I would say like Neon and Fable comparable, Astra very lazy on this task.
the web design let's look at a 3d demo so this is interactive 3d cyberpunk tokyo alley in the rain again everybody gets the same task but they these models need to build it to the best of their ability right so this is a fable okay lighting and sound scrolling walk around can zoom in zoom out interesting i like it w and s you can walk around the valley but you cannot turn around that much i like it this is very good okay let's look at astra oh oh no oh no I'm starting to see why people hate Astra.
And the ground is completely bugged. What is this? And the dragging is inverted, guys.
Obviously, you cannot see it how I drag my mouse, but this is the opposite side. Oh, my God. The rain district.
This is way worse than Fable. Like, way, way, way worse. Let's look at Neon.
Okay. So Neon doesn't have the characters, but it looks good. The reflections are good.
You can also move up and down. So it gives more creative freedom. Less detail than Fable.
But man, this is not bad. So Fable wins this one. Neon in the second and way behind Astra.
That was very embarrassing for OpenAI. What is this? Also, there are these pre -configured camera.
This is not bad from Neon. All right. So next, let's look at pure code art.
I love this. I love stuff like this. This is when you give these models a task to just do art with no image models.
And it's pure. code. So here is Patagonia mountain range at sunrise painted in pure coat here.
We have fable and how it paints it beautiful animation really good It's like the artsy vibe not really realism Astra took a different art style didn't give us an animation just a result It's not bad, but it's worse than fable. Let's look at neon neon made it animated Okay, it gave us the slider for the hours So I would say neon has the worst like art style or design But it put in more effort by adding the waves and adding the water animation and the day and night cycle.
It's debatable whether Neon and Fable wins. I think the Fable has the better artistic style.
Neon has more effort. Astra probably lagging behind again. And then we have an arcade game.
So let's look at it. Addictive one button arcade game. Let's see.
So this is Fable. Okay. That's pretty good.
See, you have a certain amount of time to click and go to the next level. It's very easy to understand. You have the aiming.
Oh, you can skip steps. Oh no. This is good.
If you release this on like your iPhone. Okay. That's crazy.
This is a bit of a cheat code where you can just like. aim to the future this is low -key good oh no yeah this is a great game okay people will be playing this forever on the phone like even like five years ago even today like i think even today if you package did well put it on the google play store put it on the ios store this could be a banger i really like this a lot of you don't know this but before my first business i was like a mobile developer you know like a game developer basically i built an android game and you know way before llms pure unity c sharp i had to write everything from scratch and learn from youtube this is a really great game okay let's look at what astra cooked up again should be one button game but the thing is it's not like obvious here as with fables so that's already slight minus points nothing major but with the fable it was immediately obvious what the game is and how to play it so let's see how streak stop the needle in the green zone hit the bullseye to build your streak okay let's see okay it's not bad the graphics i like this kind of graphic style i guess this is getting smaller and smaller over time oh i missed it it's not bad
It's less visual than the other one because that one's like moving. This one is all in the same place and you're just turning the dial. So it's definitely less engaging, less addictive, but I cannot be too harsh.
It's not bad, but worth the fable. Let's see what Neon cooked up. We got the Orbita.
Oh, is it a similar concept? There's no way. What is the rules?
Wait. Oh, you're orbiting.
Okay, you cannot hit the core. You need to orbit around. Okay, so it's similar.
Are these models like trained from the same art style? Okay, you need to dodge the core. And you need to go to these orbits.
Oh, that's crazy. That's crazy. Wait, can you like...
Oh, you gotta re -click. This is interesting mechanics. Oh, you gotta re -click.
This is interesting mechanics. Loki, I like this. Okay, this is not bad.
Yo, is this better than the Fable one? This might be more high effort than the Fable one. The Fable one was simpler.
I know I should be commentating more, but this is good. I think I need to give number one to Neon. The Fable had like better, simpler graphics.
This game is good. Oh, I'm dead. Neon number one, Fable number two, Astra definitely behind.
So this is the coding test. You can see that Neon is very, very competitive. All right, so now let's talk about where to actually use Gemini for Neon.
And probably the most native way to use it is actually anti -gravity CLI, not the Gemini CLI. I know that was a thing like nine months ago or whatever. That's pretty much dead.
not mistaking they are not even updating that so anti -gravity cli is the main product from google when it comes to cli type in anti -gravity cli boom click the first link and just install it It gives you a one -liner installer. Boom, copy that, paste it in and hit enter.
This will install it on your computer. Now, in terms of the ADE or the IDE, you know, agent development environment, basically you shouldn't be using codecs or cloud desktop or anti -gravity. It's pretty obvious that like an ADE that offers all of them, right?
Such as Cloudroom is the way to go. Using cloud code, codecs, anti -gravity, Droid, FX, Pi, Cursor, this is the way to go. This is obviously the way to build software.
So obviously we do have support inside of Cloudroom for anti -gravity CLI, and you can just use it with your subscription. So just like you're using Cloud Code, just like you're using Codex, in the same interface, you can be using Neon now with your anti -gravity subscription. So there's four reasoning efforts, extra high, high, medium, and low.
There's no max. And again, most of the time, just... Keep it on high.
It's more than good enough. It's very fast to respond. I mean, I would say maybe like a bit slower than Opus, but similar as Opus 5 .5, which is great, obviously.
You can build with it as your primary model. And in terms of the subscription, obviously Google has a bunch of AI plans. Let's look at this.
So there's three main plans. Google AI Plus, Google AI Pro, and Google AI Ultra. I would say Google AI Pro is the best.
I mean, this gives you Neon. It's $20 a month. And that's all you need, basically.
Google AI Plus, I don't know if it's useful or it's basically useless. But the benefit of Google is just they have so much software. They can give you like storage and YouTube Premium and all that stuff, right?
You can view the plan benefits and literally gives you like loads and loads of stuff. YouTube with fewer ads. Okay, this one is ad -free.
So YouTube Premium is only on the Ultra. This is like fewer ads. But yeah, you get five terabytes of storage.
Here you get 20. terabytes of storage but google ultra is quite expensive so i would recommend starting at ai pro you can just you know get a bunch of usage if you need to upgrade you can get ai ultra but the main thing is neon is here carbon is only on ultra but again neon is better than carbon so there's literally nothing to worry about so the best simplest setup for the most usage of neon is anti -gravity cli for the harness and then google ai pro for the subscription and let's look at this actually how much more ai usage you get based on the subscriptions right so this is recent data from pavel and you can see that cloth still the most subsidized subscription.
The best deal is Opus 5 .5, right? So if you want to have the best, most usage, use Opus. Fable is worse because it only gives you half of the plan.
So it's way less subsidized than Opus. Now we have Neon, which is basically the second best at 25 .8x multiple. So when you pay for the $100 Google AI Ultra, you get roughly $2 ,600 of API usage equivalent.
Remember that Neon is already very API efficient. So this is like really great multiplier. Compared to Opus and Fable.
So yeah. The next best one is Grog. With the Unified X subscription.
Which is you know. Unified with Cursor and X. Chair GPT Pro.
Total disaster. I don't know. What is this?
What are they doing? But this GPT 6 .0. So I don't know why they didn't include 6 .2.
I think the usage is actually very generous. On 6 .2. Sol and Astra.
Better than 6 .1 for sure. But hey. Look at this.
Google AI Ultra is becoming better and better subscription. By the week. So this setup is very.
Very economically viable. And I think more people are going to be running anti -gravity as their harness with Neon as their model, right? Maybe alongside Cloud Code as their primary or secondary.
It is up to you, but the deal is really solid, both on the Pro plan and on the Google AI Ultra plan. Now let's talk about reasoning effort. As I mentioned, there's four modes, low, medium, high, and extra high.
For most of these cases, just keep it to high. Like if you don't want to think about it, if you're one of these people who doesn't like to switch reasoning effort often, just keep it to Neon and high. You don't need to change it ever.
You're going to be fine. If you're doing more advanced coding, and you want it to perform the best, you can switch to extra high for those tasks that matter and just keep it to high for other tasks.
In fact, on a lot of benchmarks, you gain very little for turning to extra high and you like double the cost, right? So on the artificial analysis index, you literally get one extra point, but with 72 % extra cost and 53 % time. So again, it only matters if you really want to squeeze it to the fullest, but it's not a good deal going to extra high.
Cursor bench, again, you get 1 .7 points at 63 % cost, right? So you literally get like... 97 % performance on high compared to extra high.
And you know, it's just not worth going that extra cost. Automation Bench from Zapier, again, plus two points, almost doubled the cost. GDP .pdf at long PDFs, you literally get the worse.
It gets worse while you pay more. So yeah, going with high is the best. Extra high matters only on certain really advanced workflows.
Okay, so let me actually show you what's possible when you take Hicksfield MCP, which is probably the best MCP for... AI image, video generation, and all creative work. And you take Gemini for Neon, which is one of the best models in the world.
There's a strong argument that it might be the best. And you actually work with them together, right? So first of all, when you go to Hicksfield, again, click the first link below the video, you can just click on MCP at the top and then select whatever.
It doesn't matter. Just click on the CLI or the cloud code. It's all the same.
Like the models can figure it out. They are smart enough, right? You can literally just give them the URL and tell it to Neon to install it on your computer.
So I can just do boom. Set this up on my MacBook. That's it.
Just send them the link and the model will set it up. Here, it will probably say I already have it because I installed it before. But if you don't have it, you can just tell the model what you want and the AI agent will figure it out, right?
So there's literally no excuse to like not generate these. If you're running a business, if you need some creatives, you can just tell your agent, yo, use Hicksfield MCP and set it up. Obviously, you can still go to the website and get inspiration about what these models can do.
It's kind of insane when you show some of these like professionals, you know, the people who are experts at... AI video and AI image, what they can create when you give it like 3D rendering, Blender, FreeJS, all that stuff. So for that, what you need is a paid plan.
There are three main options for Higgs field. You can either do monthly or annual. I went with the plus.
This gives you 12. 100 credits a month. Lots of the Frontier AI models and video models.
If you need more generations than that and the best deal, then go with the Ultra. It gives you the most credits for the lowest price. And you can obviously choose like higher packages as well that have even bigger discounts.
But either way, choose a paid plan and this will unlock the... MCP as well as the Higgs field CLI so that any AI agents can basically generate images and videos just by you talking with the agent in plain English you don't have to know anything you don't have to know any terminal commands you don't have to learn the UI even it's never been easier to generate cutting edge videos and images with these AI models so this agent is setting it up it's neon it just opened a new browser tab so literally you just connect the codex plugin I already have the CLI so that's enough you can do this for codex cloud code or directly the CLI and use all of them through your agent like through the agent you already use.
Just very easy to integrate with Higgs field and get started. So again, if you don't have Higgs field, just get one of the paid plans. It's the first link below the video and you can get access to all of the AI frontier image models and video models from a single platform on a single subscription.
And I'm going to show you just how easy it is to run it with Neon creating graphics, right? We're going to run this prompt. Actually, let me send it.
So it's generating. Boom. Build an interactive 3D cyberpunk arcade page with a playable minigame.
Use Higgs field to generate a custom hero visual. A 3D prompt. and the Retro Synth Soundtrack 11 Lives Music.
Oh, that's another thing. Hicksfield also has all the music, right? So if we go to the Hicksfield website, you can see the audio has all of the main models.
See audio 1 .0, 11v4, 11v4 Turbo, v3, Quen Audio, Minimax, Seed Speech. So if you give the AI agent the Hicksfield CLI or the MCP, they can basically use all of these offerings. They can see what's available when it comes to image, video, audio, agentic, and use whatever is needed to deliver the result, right?
So you don't have to worry about, hmm, which AI... audio model am I going to use? Which AI video model am I going to use today?
Which image model should I use for my thumbnails or for my marketing and advertising? Just talk to a Frontier agent, give him the Hicksfield CLI, and the agent will take care of the rest. So we'll see what the result of this prompt is.
Okay, so I just hit my subscription here. So I upgraded to the Google AI Ultra, which is the $100 one. So let's see if this recovers.
Hopefully it does. So yeah, the limits on the Pro are not that generous. I mean, I was running quite a bit of Neon, but Ultra is definitely going to be a better deal.
But it is $100, right? So a lot of people cannot afford that. So Neon still, this is not like a deep -seek V4 .1 flash, right?
It's a frontier model. So even though it's more cost -efficient than Fable, it's still gonna run out faster than a small model. And by the way, if you have any like cursor credits or any other harness, you can just run it there as well.
So we can just start it inside of the cursor CLI. Maybe I can run it on medium so it finishes a bit faster and we can compare both of these. While this is running, let me see some of these inspirations from Higgs field.
Because there's quite a bit of level difference in terms of people who are beginners and people who are experts at AI launch videos. This is like incredible, right? That Fable 5 .2 with Higgs field access can do this.
I mean, these product reviews, UGC creators, a lot of brands spend literally thousands of dollars on this. And now it's like one prompt, you know, a couple of credits. It's incredible what's possible.
Physical products. If you're running a Shopify store, if you're doing dropshipping, if you're selling any type of merch, you know, if you have an Airbnb, if you have an apartment, you can enhance the photos with better lighting. Hook multiplier.
Different hooks for... creatives yeah i mean really imagination is your only limit like given the number of image video and audio models you have here and how easy literally frictionless it is to give to your agents through these options you're just limited by your own imagination okay so this guy has recovered and the one running in cursor is also running nice by the way here's an explanation of how it works when you run jim and i for neon and hicksfield together to build something so first you pick the idea right the agent can browse hicksfield portfolio you can use the mcp to list out all of the models available and you give it the idea Then Neon codes and does the API calls, right?
So one agent uses all of the media models. 3D plus sprites, Mesh7, HinYuan 3D, AutoSprite, Voice and Sound Effects, 11 Labs V4, 11 Labs Music V2 .5, Video and Art, CDN 2 .5. You don't even need to remember all these models.
You don't even need to know what's the best model. And then you have the real result after that. You just literally wait a few minutes and it's done.
Like, this is how people solve the Riemann hypothesis. Just telling an agent which direction to go into. Okay, so this is still running.
Let's talk about writing. A lot of people are saying that Neon might be the best writing model. This guy said...
It's by far the most human sounding non -slop writing model available today. So it's not just for code, right? This is for copywriting, articles writing, any type of creative writing.
There's plenty of other surprises below just coding. It's number one on AA briefcase, presentation quality, 45 Elo over 75 .5. It's also number one on values, finance, legal and tax.
So on many different domains, Neon is the best model in the world. It's also number one on vending bench where it has to run a vending machine business. And it even lies in negotiations with Facebook.
quotes to you know do whatever it takes to win so this model will fight for you it will be on your side when you use it also it rarely makes things up so look at this this is pretty staggering when it doesn't know neon says i don't know well whereas other frontier models are more likely to just make something up invent an answer which neon only does 14 of the time opus does it half of the time and fable and astra do it like 60 of the time which is pretty crazy if you think about it deep mind definitely figured something out here with gemini for neon all right so before we look at more 3d stuff that google published like Some wild demos.
Let's look at the result of the tasks we started. This is the medium one. Jack into the grid.
Neon grid. I'll just tell both of them to start it on localhost. And we'll see how high differs from the medium.
Okay, so this one is the medium one. Okay, let's play it. Okay, so we have the 3D model here of the arcade machine.
I see. Very nice mesh. The game is pretty basic here.
Not bad at all. Inspect 3D prop. Let's look at that.
So this is the medium. Not that impressive. Let's look at the high one.
You can see this is way more exciting already. Concerned coin play the mini game in music 60 second loop some more effort than the other one The other one was shorter loop. How do you move?
Oh, I see you click like this This is much better. The difference between medium and high is much better. That's crazy The controls are a bit weird.
It's actually good. I was playing Pressing A and D, but you can control through mouse.
That's much better. Yeah, this is smooth. This is nice.
And the music was good as well. Yeah, so you can clearly see that you should run it on high and not medium reasoning effort. So when you give it 3D tools, it's really insane.
This is Alan Turing's machine from 1936. Neon pulled the references from the college, recreated everything in Blender. Incredible.
It will be possible to revisit historical events and replay major situations and moments throughout history if you just have the right information about it. Because these... models can just fill in all the gaps this is a 3d simulator from one blank interactive 3d simulator skiing with the sound as well if you are a skier this is very good yeah it stops when it does a sharp turn this is accurate physics this is wild imagine like what it's going to look like six months from now and going straight it puts you actually into hyperspeed this is real good in the trees a little bit of off -piste and you can stop that's amazing that's actually how it works now the reason why this model is better at some things and you know claude is better at others is because training data like at this here we see that fable is really great at predicting little studied proteins and protein folding and stuff like that whereas if you see everything else like astra and carbon is quite low why because anthropic included a lot of that scientific you know biology data which is obviously like how these ai models work if you include a lot of that data in the training data the model is going to be better at that domain which is both
positive and negative it's hopeful because if you get more of that high quality data the model will keep improving and there is no signs of hitting a wall in fact the improvements are even accelerating it's negative in the fact that it's not as general right you would expect like a generally smart model to just you know an AGI the G and AGI stands for general intelligence so like you would expect that model to be able to do other stuff that wasn't in the training data but we're not seeing that as much as we should have we're seeing that like if you have a lot of this right data like you know DeepMind had a lot of video and multi -modal data you can see the Gemini models performed there was really well on the day job the finance benchmark you can see is the cheapest leading model Astra is worse and 1 .5x Neon's cost Fable is also worst.
And 14 times Neon's cost, which is just crazy. On a simple bench, it's second best after Fable. And it absolutely crushes on long document processing.
I guess Google has a lot of that data. Probably even from Google Drive, even though they wouldn't admit that. Maybe some, you know, highly masked, scrubbed off for privacy.
You can see that Gemini for Neon crushes on long documents work. Let's talk about the test of the YouTube thumbnails. So again, this is what I did with Neon, Fable, and Astra.
Again, with the Higgs field MCP. So the results are as follows. First test, the concept is Neon beats Fable 5 .2 at 1 .7 of the cost.
So each of these models have to write the prompt and they all were using Nano Banana Pro. So you can see that they're very different. I mean, there's some similarity between Fable and Neon actually.
Both are using like the 7X cheaper and the logo there. And both use my same face, but AI generated. So I don't like it.
So Astra is the best here. It's my real face, not AI generated. And it has the creativity of like the 1 .7 of cost.
Let's look at... second one google made its own flagship pointless again similarities between fable and neon very interesting in how both of them prompt the image models again both are like too saturated to mr beast style i think the astra is the most creative here which is crazy because on the previous like coding tasks astra was lazy but now it's performing really well on these thumbnails compared to the other two and last thumbnail test i did is ai coding agents plus hicksfield creative studio and you can see that now it's a bit more variety i think neon actually the worst this doesn't look like me at all i like the graphics though the graphics are probably the best of the left but it messed up the hicksfield logo that's not good both astra and fable got the hicksfield logo correctly i don't like the astra one i don't know like the car my agent shot this doesn't make sense so i think fable we have to give it to fable here because it has the correct logo agents make videos and it has my face actually you know if neon had my face it would have won here and if it corrected the hicksfield logo but yeah overall astra surprised how good it is at thumbnails
And it's surprising how close Neon and Fable are together. Maybe there's some shared training data going on. All right, let's touch on cyber because obviously every single month is getting crazier and crazier.
You can see how this is going exponential. 2026 is basically like... The past, I don't know, four years combined in terms of major software vulnerabilities.
And Neon is a top cyber model. So at webhacking, aka the benchmark web exploit bench, Neon has 51 .3 versus Opus only having 37 .1. Finding bugs, SEC Bench Pro is number one at 90%.
Finding bugs at CVE Bench, Fable is number one. Neon is number two, like 4 % behind. Probably a lot cheaper.
So this is a great model. So the way I use it is inside of loops. So I just have a bunch of loops set up.
where the models alternate. So for example, here, these models are alternating, right? Between Opus 5 .5, GPT 6 .0.
What I'm going to do is I'm going to add Neon to a lot of these. So it runs like one day Neon, one day Opus, one day Fable, one day Astra, one day Sol. So you have a higher variety in terms of which models are running.
And each one will spot something different on the code base, right? So each frontier model, you should run it at least once a week on all your repos. They literally, I don't care what you're building.
This is just a cybersecurity essential when we're... during the post -AGI era. And Neon should be one of your top favorites.
With Opus 5 .5 and Neon being like very similar, both cost -effective, I think you should run it probably as your main, honestly. On net, it's better than Opus at cybersecurity, and it's obviously a lot cheaper than Fable. So if you're going to run these loops, and again, I posted all of them as templates here, you can literally one -click add them to your repos.
So if you have Cloud Room, just go to loops. top left, click on templates, and start running these yourself on your repositories so that you protect yourself and your code base. Because this is how cybersecurity works, where both the attackers and the defenders are running swarms of agents.
So the only way you can protect yourself against a hacker with a swarm of agents is to run frontier models and obliterated models on your own code bases. Models like Neon, just on a loop, trying to find bugs, trying to find security vulnerabilities by themselves, and then patching them by themselves. Agent loops are absolute game changer.
If you're still sleeping on them, log in and start using them. So basically every night it would run, it would scan the repost vulnerabilities. If it doesn't find anything, it self -archives, right?
That's the beautiful part. It doesn't bother you unless it finds something serious, something actually worth fixing. Then it can prove it, right?
It can check it with a second frontier model from a different harness, then build a simple fix. The other model reviews it and it pushes by itself. Or if you want to be in the loop, you can review it yourself.
Depends how aggressive you want to be when it comes to using these AI agents. Now, let me show you how to run this in the terminal because that's probably the simplest way when you install it. So just open the terminal.
type in a g y not i a g y that's the shortcut for anti -gravity cli you can select the color whatever google wants to ask you if you want to support them and you know by allowing google to collect interaction data so you can hit enter to toggle if you don't want to enable that done you trust this folder yes i do trust this folder okay so then you can do shift tab to switch between auto plan and accept edit modes i don't know why they still have plan mode everybody is removing it but for some reason anti -gravity still has it so i'm just going to go into auto then reasoning effort you can do slash model okay you can do the effort here yeah so low medium high extra high i'm gonna go here neon let's send it high do you have the hicks field mcp see when it's working it doesn't show as much reasoning as i saw in cloud room like here you can inspect a lot of the tool calls it did i don't know why it's not showing in the terminal okay let me interrupt it i'm gonna switch to do you see hicks will see a lot on my computer okay so i don't know why it switched to plan mode even though i put it to auto mode yes it's installed okay so now say
Create a 15 second video with the best AI video models on Hicksfield about the history of AI industry. Let's see what this can cook up. Okay, so it created a plan.
This is why the anti -gravity harness is like less optimized than Clodder Codex because I didn't say anything about a plan. It still created a plan. I don't know.
When I was using it in Cloudroom, it didn't do these things, perhaps because the default is, you know, like full access and we just have a lot of optimizations that don't have it overthink like this. But yeah, I'm going to say yes. 16 to 9 audio, up to you, up to you into downloads.
Don't overthink, just create this with Xfield. All right, so it's finally finished. It took some time.
I don't know, like five plus minutes, I feel like. So it wasn't the fastest. I think Opus would have done it faster.
But keep in mind also the video generation itself takes some time. So this is the result right here. Let's watch it.
The history of AI.
Okay, so even though it's 15 seconds, you can see there is some sort of a storyline or plot. We start in the 50s with very ancient computers. Then we move to the modern era with the Nvidia racks in a data center.
And then I presume like some period 20 years from now, I don't know, who knows if it will look like this.
And then singularity, man, god means artificial god. Crazy. Yeah, I mean, not bad.
I like the music as well. Not bad at all, given that we ran it on Neon Medium. Okay, this is another test I did.
Again, comparing Neon to Fable and Astra. So, let's look at it.
So, by the way, I told them. to create a launch video for Cloudroom, right? So each of them had the same task, just context of going to the website, literally that's it.
So they could, I guess, scroll it, they could do whatever. And it's up to that agent to figure out how to create the launch video, right? Using Higgs field MCP, obviously.
So let's look at the result. This is Fable first. Close the laptop.
Your agents keep shipping.
Not bad. I mean the voice like The video is good. The voice is a bit It's obviously AI Close the laptop.
Yeah, I just This is not a good place for the voice. I mean, but the video is really good. It's crazy This is okay -ish, you know, it's not really how the UI looks like inside of the app keep shipping yeah sound effects are good video is good just the voice is at a weird moment now astra your laptop sleeps you're so used the same voice which again it just doesn't feel right but very different like close the laptop is the same but then very different art style Your laptop sleeps your coding agents keep going Okay now neon I would say a fable was better You're living in the cloud like this doesn't feel like a cloud made of rocks, I don't know And neon is very different No voice.
So I would say Neon has to be my favorite here. First of all, the creative uniqueness is very different than what Astra and Fable decided to go with. And also no annoying voice.
I don't know. I think a different voice model would have probably been better. Or a few iterations of the model.
But yeah, you can see that Neon is very, very capable. Both at coding, as I showed earlier, with the games and 3D. Images and videos as well.
So here are my honest thoughts on neon Right now it's probably the best value like workhorse model, right? It's like the it's like opus. It's a bit slower, but I think it's more powerful and Yeah, I mean on a lot of benchmarks.
It's completely demolishes opus Astra was winning in the creative direction and thumbnails in our tests So Astra is very weird like on the coding test. It was very lazy, but then on some of the thumbnails it was by far the best so i don't know what's what's up with open ai but right now it really seems that anthropic and google are pulling ahead over things short tasks for sure opus has faster ttft time to first token so if you need something done very quickly like 10 20 seconds use opus 5 .5 medium or high um there is no safety report which i've seen Basically, I think what happened is that Google was pushing a bit, not overthinking the safety, just trying to catch up, which I agree with, by the way.
It's just that some people on Twitter found interesting ways to use Neon that other frontier models would probably reject. And very low rates of cheating, which is good. I think that's actually the alignment.
The model doesn't lie like other models. And obviously there's the full model card from DeepMind, so you can read that or have your agent read that. So yeah, overall really powerful model.
I would highly recommend you guys use it as your daily. Try it, don't just be addicted to Opus 5 .5. By the way, if you're not using Neon or Opus 5 .5, what are you doing?
These are definitely the two best models right now to use as your daily driver, as your main model. And then, you know, obviously if you need to, you can... use some more Fable for that extra power or like a big model smell.
But Neon is literally outperforming Fable on a lot of benchmarks. So use Neon or Opus 5 .5 as your primarily daily model right now. Anything else is kind of trolling to be fair.
That said, if you don't have Hicksfield, what are you doing? This is the single platform for all AI image, AI video, AI voice. You can see how easy it is.
Literally, I gave Neon... the CLI both inside of Cloudroom and we should do it in the terminal as well and just you talk to it in plain English and can use all of the different models and tools that Hicksfield has to offer through the MCP so if you don't have Hicksfield yet get one of the paid plans right now they have some crazy discounts for Black Friday you can see that up to 63 % off so it's literally the best time to buy so click the first link below the video get either the plus plan or ultra plan whatever makes sense for you and start using My favorite way to use Hicksfield is through the MCP, but obviously you can use it directly in the website as well.
With that being said, thank you guys for watching and I wish you a wonderful productive week. See ya!
The Hook

The bait, then the rug-pull.

Google's mid-tier Gemini 4 Neon arrived six days after the previous model and immediately topped CursorBench, beating its own bigger sibling and Fable 5.2 at a fraction of the price. This breakdown walks through the benchmarks, five rounds of live head-to-head tests, the plan and CLI to run it on, and the effort setting that saves the most money.

Frameworks

Named ideas worth stealing.

48:20model

Daily driver plus heavy hitter

  1. Neon or Opus 5.5 as the default model
  2. Fable 5.2 only when a task needs extra power
  3. Opus 5.5 for 10 to 20 second quick tasks

Split work by cost and speed instead of always reaching for the biggest model.

Steal fordeciding a team-wide default model
26:37list

Effort ladder

  1. Low
  2. Medium
  3. High (default)
  4. Extra high (hardest coding only)

High gets about 97% of extra-high performance at much lower cost; medium visibly underdelivers on full builds.

Steal forsetting agent defaults
39:32list

Nightly security loop

  1. Every night a loop starts a frontier model
  2. Find: scan the repo for vulnerabilities
  3. Prove it: reproduce the bug, drop false alarms
  4. Fix: write the patch and a test
  5. Review: a second model or a human checks, then push
  6. Self-archive if nothing real was found

Rotate models from different labs across nights so each one catches different issues.

Steal forany team shipping AI-written code
24:27list

Cheapest strong setup for Neon

  1. Harness: Antigravity CLI
  2. Plan: Google AI Pro, $20 a month
  3. Effort: high
  4. Upgrade to Ultra ($100) only if you hit limits

The simplest path to the most Neon usage per dollar.

Steal foradding a second lab to an AI stack
CTA Breakdown

How they asked for the click.

VERBAL ASK
49:35link
“Click the first link below the video, get either the plus plan or ultra plan whatever makes sense for you.”

Sponsor read placed three times (mid-image test, a full build demo, and the outro) with a Black Friday discount for urgency. The mid-video build demo doubles as product proof.

Storyboard

Visual structure at a glance.

claim
hookclaim00:00
CursorBench
promiseCursorBench02:00
release cadence
valuerelease cadence03:26
image test
valueimage test07:54
why Google wins
valuewhy Google wins12:11
coding showdown
valuecoding showdown17:33
Antigravity CLI
valueAntigravity CLI23:21
plan multipliers
valueplan multipliers25:56
use high
valueuse high26:57
beyond coding
valuebeyond coding33:26
thumbnails
valuethumbnails38:19
security loop
valuesecurity loop40:56
verdict
ctaverdict47:48
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.