Why Modern Creator?

Nick Saraev · YouTube

I Tested Kimi K3 So You Don't Have To

A live side-by-side of Moonshot's open-source Kimi K3 against Anthropic's Fable 5 across coding benchmarks, generative design, 3D scenes, and mini-games — followed by a case for why the real bottleneck is about to stop being intelligence at all.

Posted

July 16th

3 months ago

Duration

09:52

Format

Review

hype

Views

40.6K

1.4K likes

Big Idea

The argument in one line.

Kimi K3 matches or beats closed frontier models like Fable 5 at a fraction of the price, which means the bottleneck on what AI can do is shifting from model intelligence to who has access to run it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're deciding whether to build on an open-source model like Kimi K3 instead of a closed vendor like Anthropic or OpenAI.
You run or advise a company evaluating a move off locked-vendor AI infrastructure toward cheaper open alternatives.
You want a fast read on how a new frontier-tier open model actually performs against benchmarks you don't have time to dig into yourself.
You're interested in the macro argument that AI intelligence is deflating fast and want the economic reasoning, not just the benchmark numbers.

SKIP IF…

You want a rigorous, controlled benchmark methodology — this is one person's one-shot demos and reactions, not a scientific comparison.
You're looking for hands-on setup instructions for running Kimi K3 yourself — none are given.

TL;DR

The full version, fast.

Moonshot's Kimi K3 dropped and lands in frontier territory on coding benchmarks, beating GPT-5.6 on software engineering and edging out Fable 5 on some terminal and program benchmarks, all while costing roughly Sonnet-5-tier prices instead of Fable-5-tier prices. Live one-shot tests across a research-report design, three WebGL scenes (an orbital traffic grid, a procedural flyover, and a particle galaxy), and two browser mini-games show Kimi K3 matching or beating Fable 5 on most, though it runs about three times slower on the heaviest scene. The larger argument: as frontier intelligence gets 30-40x cheaper per year and lands in open-source models anyone can run, the bottleneck on what gets built stops being model quality and starts being who has the distribution and compute to deploy it at scale.

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 – 00:44

01 · Kimi K3 drops

Cold open: Moonshot's Kimi K3 releases and immediately lands near frontier-model territory on software-engineering benchmarks, upsetting the closed-source balance of power.

00:44 – 01:45

02 · The benchmarks

Kimi K3 beats GPT-5.6 on frontier software engineering by Moonshot's internal numbers, and outperforms Fable 5 on Terminal Bench 2.1 and Program Bench, all at roughly Sonnet-5 pricing instead of Fable-5 pricing.

01:45 – 02:10

03 · Why this matters to real companies

Clients and Maker School members are already asking Nick what it would take to move billion-dollar infrastructure off closed vendors like OpenAI and Anthropic onto Kimi K3.

02:10 – 03:06

04 · Research report design test

Same prompt, one shot, both models: a 'State of Remote Work in 2026' report. Kimi K3 and Fable 5 produce near-identical layouts, suggesting distillation or dataset overlap between the two.

03:06 – 04:01

05 · 3D scene: Orbital Traffic Grid

The first WebGL 3D scene is where quality clearly diverges — Kimi K3's version is described as not even comparable to Fable 5's, with a much richer, more atmospheric result.

04:01 – 04:29

06 · 3D scene: procedural flyover

An endless procedurally-generated landscape flyover; Kimi K3's landscape reads cleaner and more polished, though its on-screen text renders with font issues.

04:29 – 05:30

07 · 3D scene: particle galaxy

A particle-based nebula/galaxy demo. Fable 5's version mostly whites out the screen; Kimi K3's version adds working sliders for star density, arm count, and spin — Kimi wins, but took roughly 3x longer to generate.

05:30 – 06:14

08 · Mini-games: Stack + physics

Both models one-shot a block-stacking mini-game; results are close in quality, framed as a straight tradeoff between waiting longer for Kimi K3's cheaper output versus paying more for Fable 5's faster one.

06:14 – 07:22

09 · Intelligence is deflating

Pivot to the macro argument: AI intelligence cost is dropping roughly 30-40x per year, and frontier-level capability will eventually run locally on cheap hardware.

07:22 – 09:51

10 · Distribution, not intelligence, is the real bottleneck

Closing monologue: once frontier intelligence is cheap and open, human knowledge work stops being the bottleneck to building things, and the constraint shifts to who has the compute and distribution to deploy models at scale — with open, non-monopolized access to frontier AI framed as the healthier outcome.

Atomic Insights

Lines worth screenshotting.

Kimi K3 landed almost band-for-band with frontier closed models on software-engineering benchmarks within hours of release.
Kimi K3 beat GPT-5.6 on frontier software-engineering benchmarks according to Moonshot's own internal numbers.
On Terminal Bench 2.1 and Program Bench, Kimi K3 outperformed Fable 5, a closed model priced far higher.
Kimi K3 runs at roughly Sonnet-5 pricing while delivering intelligence close to Fable-5 tier, upending the value calculus for anyone paying frontier prices.
On identical one-shot prompts, Kimi K3 and Fable 5 produced near-identical page layouts, suggesting Kimi K3 may have distilled from Fable 5 outputs or shares training-data overlap.
On a 3D orbital traffic grid scene, Kimi K3's output was described as not even comparable to Fable 5's — a clear quality gap in Kimi K3's favor.
Kimi K3's particle-galaxy demo included adjustable sliders for star density, arm count, and spin that the creator could not replicate in Fable 5's version.
The heaviest 3D scene took Kimi K3 roughly three times longer to generate than Fable 5, so the quality edge came with a real speed cost.
AI intelligence is dropping in cost by an estimated 30 to 40 times per year, pound for pound.
The creator argues that once frontier intelligence is cheap and open-source, the bottleneck for what gets built shifts from model capability to distribution and compute access.
If a lab open-sources a frontier-tier model the way Moonshot did with Kimi K2 and now K3, it materially changes the economic calculus for closed labs like Anthropic and OpenAI.
The creator frames a thought experiment: even if model intelligence stopped improving today, humans would spend the next 20-30 years just weaving current-level AI into every layer of society.
The core argument is that human knowledge work — not model intelligence — has been the bottleneck to building anything, and cheap open models remove that bottleneck for anyone with GPU access.

Takeaway

Frontier AI is deflating faster than you're pricing it

WHAT TO LEARN

A new open-source model can land in frontier territory within hours of release, so any bet on locked-in AI pricing should assume the cost floor keeps falling out from under it.

A new open-source model landed within striking distance of closed frontier models on coding benchmarks the same day it released, which means the gap between 'open and cheap' and 'closed and expensive' can close almost overnight.
Benchmarks alone don't tell the full story — the same prompt produced near-identical output from two different models on a simple design task, but a clear quality gap on a demanding 3D scene, so capability differences show up unevenly across task types.
A cheaper, slower model and a pricier, faster one can land at genuinely comparable output quality, turning the choice into a real tradeoff between wait time and cost rather than a clear winner.
If AI intelligence keeps getting dramatically cheaper every year, today's expensive frontier capability becomes tomorrow's near-free commodity, which changes the math on any long-term bet built around current AI pricing.
As raw model capability becomes cheap and widely accessible, the limiting factor on what actually gets built shifts from 'is the model good enough' to 'who has the compute and reach to put it to work.'
When frontier capability is available in open form rather than locked behind a small number of vendors, the resulting competition is framed as healthier for buyers than a market where intelligence is concentrated in one country or one company.

Glossary

Terms worth knowing.

Kimi K3: An open-source large language model from Chinese AI lab Moonshot, released as a successor to Kimi K2, positioned as a frontier-tier coding and reasoning model.
Moonshot: The Chinese AI lab that develops the Kimi family of models and has released prior versions as open source on platforms like Ollama.
Terminal Bench: A benchmark that scores how well a model can operate a command-line terminal to complete tasks, used as a proxy for real developer-tool usefulness.
One-shot: Generating a complete result from a single prompt with no follow-up corrections or iteration, used here to compare raw model capability without prompt engineering.
Distillation: Training a smaller or newer model by learning from the outputs of a larger existing model, which can produce visually or stylistically similar results between the two.

Resources

Things they pointed at.

00:18toolKimi K3 (Moonshot)

00:18toolKimi K2 (prior Moonshot release)

00:22toolOllama

00:44toolGPT-5.6

00:44toolFable 5 (Anthropic)

01:12toolSonnet 5 (Anthropic)

01:45communityMaker School

Quotables

Lines you could clip.

00:44

“Frontier software engineering, it actually whoops GPT 5.6's ass.”

blunt, punchy claim with a specific competitor named→ TikTok hook↗ Tweet quote

01:12

“Why would you pay Fable prices when you can get, you know, 98% now of Fable five intelligence at Sonnet five costs?”

distills the entire value argument into one line→ IG reel cold open↗ Tweet quote

06:26

“The intelligence isn't getting more expensive... it's dropping something like 30 to 40 times a year based off of my calculation.”

concrete, quotable stat on AI cost trends→ newsletter pull-quote↗ Tweet quote

08:47

“Human knowledge work is the bottleneck to virtually everything... What does the world do? How do we value economic and human economic labor specifically?”

opens the video's central philosophical question→ TikTok hook↗ Tweet quote

08:47

“The model intelligence really isn't the bottleneck anymore, if that makes sense. It's the distribution.”

the thesis line of the whole back half of the video→ IG reel cold open↗ 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.

metaphoranalogystory

Boy, oh, boy, oh, boy, has a lot changed in the last few hours. Kimmy k three dropped, and it has completely upset both the balance of power with closed source models, as well as the balance of power between the American sort of hegemony model developers and the Chinese.

So this model company, Moonshot, has been working on Kimi for a while now. And I've talked about it a little bit on the channel, you know, they've released them open source on platforms like OLAMA. But this really takes the cake.

I mean, as you see here, when they dropped just a few hours ago, Kimi k three is almost band for band on deep software engineering. I'd very confidently say this is now in frontier model territory. Frontier software engineering, it actually whoops GPT 5.6 souls ass.

Their own internal benchmarks, as biased as they may be, tell a similar story. And a few of these terminal style benchmarks, like terminal bench 2.1, show it better even than Fable.

In this case, program bench far better than Fable. So this is gonna change the, you know, balance of user preferences quite a bit as well. Because why would you pay Fable prices when you can get, you know, 98% now of Fable five intelligence at Sonnet five costs?

And despite how most American companies and wider companies feel about giving a bunch of data to, you know, quote unquote scary Chinese models, there's a lot to be said. I mean, if they pump this thing out into the open source ecosystem, like they did with Kimik 2.7 code, that's significantly going to change the economic calculus of something like Anthropic, OpenAI, and so on.

So like, I'm gonna try these models out right now with you guys, so you could see just what it means when we're not staring at isolated benchmarks. But I did wanna be clear, this is a major upset. Like a lot of people in maker school, a lot of people in the companies that I consult with, they've already asked me stuff like, Nick, what would it take to move my entire infrastructure over to Kimi k three versus continue operating it on, you know, closed sort of locked vendors like OpenAI and Anthropic.

And some of these are billion dollar companies. I mean, this is not a small feat. Alright.

So first, we have some pretty basic design. This is a simple state of remote work, sort of a research paper. As you could see, the designs now are pretty similar.

And there's a couple of reasons for that. Odds are Kimi k three distilled from Fable five directly, or there was a fair amount of sort of like cross contamination of datasets and so on and so forth going on.

I think they're also doing a lot of like opinionated training to try and get the models to do things like vary font styles and so on and so forth. So nothing really exciting here. Obviously, the quality of Kimi k three is gonna be pretty similar to that of Fable.

You can see the same thing happening here on like this little electric vehicle one. Fable decided to use sort of a bland research report for a sleep review, whereas Kimmy over here took a little bit more creative of a license.

But you can see that they still did something very similar. They used sort of like this old school book style first letter, really, really big.

I'd actually say Kimmy probably sized that better, which is interesting. Okay. Now we have some three d and WebGL scenes, and this is where the qualities really start to diverge.

So on the right hand side here, this is sort of like an orbital traffic grid designed by Fable. So you guys can see, you know, it did it did the job. It is technically a globe and it has information sort of transferring from one side to the next.

It's kind of neat. However, if you compare that with Kimi's, you'll find a world of difference. I mean, this is just not even comparable, not even close.

I can still zoom in. I can still do more or less everything, but, you know, we have this extremely like, I don't know, otherworldly sort of vibe.

It's also a whole country with like information highways clearly spread throughout. And as I zoom in, things get a little bit clearer.

You can see it's a high quality, high resolution map with this overlaid. I can't say the same for this. So, you know, keep in mind, these are one shots obviously, but I would say this did on par if not far better.

Alright. Now this is a Fable five endless flyover of some map that's being procedurally generated on the right hand side.

And, you know, like, it's pretty cool. There's some interesting rumbling going on, which kind of scares me if I'm being honest. You know, I can do the same thing with Kimmy k three over here.

I'd say the landscape is a lot cleaner and a lot prettier, probably a lot more beautiful. That said, there's some issues with reading the font up top. I can change the speed by scrolling around.

I'm not going to scroll too much just because I want to show you guys more. If I scroll down here, we're now sort of doing a particle style galaxy, a nebula. And on the right hand side here is fables.

As you could see, most of the screen just turns white when I go down here. I do have the ability to move this. This kind of reminds me of Spore.

I don't know if you guys ever used to play Spore, but I used to play a lot of that. On the right hand side here, looks like we have the ability to adjust the slider and, you know, the the brightness of the stars and stuff like that myself, which is kind of cool. If I click on this plus button, you can see I can change the the amount of star dust, the number of arms, you know, the spin, and so on and so forth, the the the chaos marker, which is kinda cool.

So if I exit that out, I mean, like, this is this is really nice. I don't know if I could do that here. I don't think I can on Fable's version.

So, yeah, I mean, the aura definitely kicked Nebula and Draste's ass on that one. I'm gonna give Kimmy k three the win, but worth noting that did take three times the amount of time that Fable did. Okay.

And then we have a couple more mini games. On the right, we have Fable five with, you know, the stacker. So this is me stacking a bunch.

Looks pretty good. And obviously, I'm not too hot at that game, which is how I lost. Over here, just gotta close some tabs before my computer has a hernia.

You know, we could see the stacking is like three-dimensional, which is kinda cool. I like that.

I would say this is far more akin to the sorts of things you'll see at, a I don't know, like, one of those casinos or rec rooms, which is kinda neat. Not really seeing too many differences here in the quality. Obviously, one model's a little bit better at something.

Another model's a little bit better at something else. The point is, one is at Sonnet five pricing, and the other one is at Fable five pricing. So sort of like choose your own minigame.

Would I rather spend two times longer for, you know, a third or a fifth of the price? Would rather get it done in eighty seconds as opposed to two hundred forty seconds, but sort of burn a hole in my wallet. What does the world look like when we all have access to Fable five intelligence in our pockets for virtually free?

Because if it's not clear, that's where all of this stuff is going. The intelligence isn't getting more expensive. Bar for bar, pound for pound, it's dropping something like 30 to 40 times a year based off of my calculation.

It is getting so overwhelmingly expensive that I see it feasible at some point in the future that you actually have large language models running on headphones capable of doing the vast majority of the lifting that Fable five is doing. And that's just in a few years. So just for the sake of argument, what do you think would happen in a world like that?

Now, don't think it takes a neuroscientist to explain this, but I have a published paper in neuroscience, so I'll try and tell you. To make a long story short, we'll have significantly democratized access to an average human intelligence.

We'll be able to spin up dozens, if not hundreds of these, literally at the snap of a finger to help us with any intellectually valuable task. When this occurs, the entire scarcity aspect of the economy will change completely. I mean, now, human knowledge work is the bottleneck to virtually everything.

You know, if you want a service, you want a website, if you want a consulting report, if you wanna develop a new piece of technology, if you wanna do material science, push the frontier of science and so on and so forth forward. You need people. Right?

Like we're people. It's our meat brains that are doing all of this. What does the world do?

How do we value economic and human economic labor specifically? What happens to the vast majority of jobs that we all have?

So an interesting thought experiment for you is, let's say for whatever reason, maybe just the laws of physics, models stop growing smarter. They just get no smarter starting right now.

So basically, we capped out at, you know, a fable five consumer level model, and we just can't grow it smarter. What do you think happens to the world over the course of the next twenty or thirty years?

Does all scientific advancement and progression and development stop? Even though the models themselves are no longer going to grow any more intelligent, human beings will come up with ways to use these models to incorporate them into society, to weave them into the fabric of social life tremendously over the course of the next twenty or thirty years.

I mean, we'll have AI assistants waking us up in the morning. We'll have AI assistants coordinating the the building and architecture of new freaking planets. We could have a Fable five level intelligence shepherd thousands of humans across a freaking generation ship in the year 2982.

The model intelligence really isn't the bottleneck anymore, if that makes sense. It's the distribution. So what Kimi here, a Kimi k three effectively solves is that distribution.

You could have everybody on planet Earth, you know, who has the GPU access, which is another story entirely. But everybody on planet Earth running this on a future, you know, Olama style setup where it's totally open. And to be clear, I don't have the answers to these questions.

I'm just asking them because I think now is the time to ask them, not in ten years when those answers could have helped us ten years ago. Right? So seeing cool model technologies like Kimi k three really fires me up.

And I'm happy because I don't want like a monopolar, a monopolar sort of style future where all of the intelligence is congregated and segregated in one country. I think we should naturally have some form of competition.

Anyway, think about it. We're all gonna have Kimi k three style models on our phones. They're all gonna be a 100% accessible, crazy uptime.

What more do you really need at our current level? And if the answer to that question is nothing, then clearly something about our current level is going to have to change. There's just no way we can sustain this current economy in a future where models do vast majority of everything.

The Hook

The bait, then the rug-pull.

Within hours of release, Kimi K3 landed close enough to frontier-tier closed models to upset the pricing power of every major AI lab — the creator spends the first half of the video proving it live, then spends the second half arguing what happens once that pattern repeats forever.

Frameworks

Named ideas worth stealing.

10:22concept

Intelligence deflation

The claim that frontier-level AI capability is getting 30-40x cheaper per year, meaning today's expensive frontier model becomes tomorrow's free commodity.

Steal forframing any pitch around why acting now on AI-dependent infrastructure beats waiting for prices to drop further

15:47concept

Distribution over intelligence

Once model intelligence is cheap and widely available, the constraint on what gets built shifts from 'is the model smart enough' to 'who has the compute and reach to deploy it.'

Steal forpositioning an open-source or self-hosted AI product against a closed competitor on access rather than raw capability

CTA Breakdown

How they asked for the click.

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.

OTHER LINKSAlso linked in the description.

Frame Gallery

Visual moments.

Frame at 00:00 from I Tested Kimi K3 So You Don't Have To

Frame at 00:13 from I Tested Kimi K3 So You Don't Have To

Frame at 00:17 from I Tested Kimi K3 So You Don't Have To

Frame at 00:24 from I Tested Kimi K3 So You Don't Have To

Frame at 00:29 from I Tested Kimi K3 So You Don't Have To

Frame at 00:40 from I Tested Kimi K3 So You Don't Have To

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.

Video of the Day14:34

Nick Saraev · Tutorial

GLM-5.2 is Basically Opus (For 1/5 the Price)

A 14-minute benchmark rebellion: seven live side-by-side demos, one OpenRouter API key, and a four-path procurement map that makes Opus 4.8 look expensive.

June 19th

23:59

Nick Saraev · Talking Head

Here's What I'd Learn Instead of AI Automation in 2027

An AI automation agency owner making seven figures a year argues the easy money in his own industry is closing fast, then lays out what's actually worth building toward next.

October 5th

36:15

Nick Saraev · Essay

The Last AI Roadmap Video You'll Ever Watch

A one-week curriculum that starts at the 1950s perceptron and ends at selling AI work, with the full reading list attached.

October 16th

07:00

Nick Saraev · Listicle

How to Use AI Like A Top 0.1% CEO

Nick Saraev breaks down seven concrete habit swaps that separate the highest-output AI users from everyone else: talk instead of type, keep AI listening through every meeting, show instead of tell, let AI interview you, demand flaws instead of opinions, and just try the thing yourself before assuming it's impossible.

October 14th

07:30

Nick Saraev · Tutorial

I Let Claude Code Run My $1M/Year Business for a Whole Week

Nick Saraev spent a week letting Claude Code audit, then automate, nearly every non-client-facing task inside his seven-figure agency.

October 13th

10:46

Nick Saraev · Review

Reverse Engineer Anything: Testing the Viral AI Decompiler Everyone's Hyping

Nick Saraev races a vanilla Claude agent against REA, the trending GitHub tool for AI-powered app cloning. He finds the underlying tech is decades old, and the agents themselves are already commoditizing the software built around them.

October 10th