Modern Creator
Why Modern Creator?
Chase AI · YouTube

Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King

Kimi K3's benchmark charts and rock-bottom per-token price look like a knockout blow to Claude and GPT — until a blind three-way build test and a real cost-per-task tally tell a much closer story.

Posted
yesterday
Duration
Format
Review
educational
Views
1.1K
56 likes
Part of the collectionThe Fable 5 PlaybookAll 45 Fable 5 breakdowns, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

Kimi K3's per-token price looks like a bargain next to Claude and GPT, but because it burns far more tokens and takes far longer to finish the same task, its real-world cost advantage nearly disappears against token-efficient frontier models.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You're deciding which LLM to route coding work through and want to weigh sticker price against the actual cost of finishing a real task.
  • You use OpenRouter or similar multi-model billing and want to understand why 'cheaper per token' doesn't always mean a cheaper bill.
  • You're curious whether open-weight Chinese models like Kimi K3 are a legitimate production alternative to Claude or GPT, not just a benchmark-chart winner.
SKIP IF…
  • You're only looking for a design/aesthetics critique of the three demo builds — the video's real value is the cost and time math, not UI polish.
  • You want guidance on running an open-weight model locally — the video is explicit that Kimi K3 needs enterprise-scale hardware, not a personal machine.
TL;DR

The full version, fast.

A new open-weight Chinese model, Kimi K3, has been going viral on benchmark charts showing it beating Fable 5 and GPT-5.6 on coding tasks while costing a fraction of the price. A blind three-way build test — a 3D globe travel dashboard — shows Kimi K3 producing solid but not best-in-class output, ranking behind Fable 5 on subjective quality. The real story is cost: Kimi K3's cheap per-token price is offset by heavy token usage and slow generation, so on a true cost-per-completed-task basis it barely beats efficient frontier models like GPT-5.6, and it is dramatically slower than either. The takeaway is to judge open-weight models on total cost and time to finish real work, not on advertised per-token pricing.

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 – 00:46

01 · Intro — the claim

States the viral claim (Kimi K3 beating Fable 5 and GPT-5.6) and previews the plan: read the benchmarks, then run a blind head-to-head frontend build test.

00:46 – 02:13

02 · The Numbers

Walks through Kimi K3's 2.8T-parameter open-weight release, its SWE-bench/coding benchmark wins, and its #1 ranking on the Frontend Code Arena blind-vote leaderboard.

02:13 – 04:34

03 · The real pricing math

Compares sticker per-token pricing (Kimi $3/$15 vs Fable $10/$50 vs GPT-5.6 $5/$30), then pivots to Artificial Analysis's cost-per-task index, which shows the price gap against GPT-5.6 shrinking to roughly 10%.

04:34 – 05:38

04 · The hallucination benchmark

Introduces the AA Omniscience Index; Fable 5 scores far ahead (40/100) while GPT-5.6 (22) and Kimi K3 (18) are close together at the bottom.

05:38 – 07:23

05 · Head to Head: Kimi K3 builds

Kimi K3, run inside Claude Code, builds a 3D globe travel dashboard ('Meridian') from an intentionally vague creative prompt. Functional but lower visual fidelity.

07:23 – 08:59

06 · Head to Head: Fable 5 builds

Fable 5 builds the same dashboard with cleaner graphics, a working day/night toggle that dynamically updates fare pricing, and higher overall polish.

08:59 – 10:00

07 · Head to Head: GPT-5.6 builds

GPT-5.6 (via Codex) goes for a minimalist sci-fi look but ships small, hard-to-read overlapping text and a weaker day/night effect; ranked third.

10:00 – 12:19

08 · Tallying tokens, time, and cost

Whiteboard tally: Kimi K3 used 21.5M tokens / 93 minutes / $8.66; Fable 5 used 3.5M tokens / 17 minutes / $11.64; Codex used 5.6M tokens / 25 minutes / $5.66.

12:19 – 13:40

09 · The Verdict

Kimi K3 is a genuinely competitive open-weight model, but its price advantage is overstated against efficient frontier models and its slow generation speed is a real practical cost.

Atomic Insights

Lines worth screenshotting.

  • Kimi K3 lists at $3 per million input tokens versus Fable 5's $10 — a 70% sticker discount that shrinks dramatically once actual token usage is counted.
  • On Artificial Analysis's cost-per-completed-task index, Kimi K3 cost 95 cents versus GPT-5.6's $1.04 — a roughly 10% gap, nowhere near the 3x difference the per-token sticker prices imply.
  • Some smaller, more token-efficient models in GPT-5.6's tier complete tasks at half of Kimi K3's true cost, despite a higher advertised per-token rate.
  • Kimi K3 needed 21.5 million tokens and 1 hour 33 minutes to build a 3D globe dashboard that Fable 5 completed in 3.5 million tokens and 17 minutes.
  • Despite burning six times more tokens than Fable 5, Kimi K3's total build cost ($8.66) still undercut Fable 5's ($11.64) — its rock-bottom per-token price outweighed its inefficiency.
  • On a hallucination-resistance benchmark (AA Omniscience Index), Fable 5 scored 40 out of 100, GPT-5.6 scored 22, and Kimi K3 scored 18 — nearly a tie between the two lowest scorers.
  • Kimi K3 ranked #1 on the Frontend Code Arena's blind head-to-head voting, but blind-vote leaderboards reflect taste as much as capability and say nothing about cost or speed.
  • In a blind three-way build test, the reviewer ranked the actual output quality Fable 5 first, Kimi K3 a close second, and GPT-5.6 third.
  • Speed, not price, is Kimi K3's biggest practical weakness — a 90-minute wait for a single build is a dealbreaker for iterative development regardless of dollar savings.
  • A model's advertised per-token price tells you almost nothing about a finished project's real cost until you know how many tokens it actually burns to get the job done.
Takeaway

Cheap per-token pricing can hide an expensive bill

COST LITERACY

A model's advertised price per million tokens tells you almost nothing about what a finished task will actually cost until you know how many tokens it burns and how long it takes.

02The Numbers
  • Benchmark leaderboards like SWE-bench and the Frontend Code Arena reward raw task performance, but a top spot doesn't tell you what a real project will cost or how long it'll take.
  • Blind A/B leaderboards are inherently subjective — voters pick a winner without knowing model identity, so the ranking reflects taste as much as capability.
03The real pricing math
  • Per-token sticker price only tells part of the story — total cost depends on how many tokens a task actually consumes, which varies wildly between models.
  • On a true cost-per-completed-task basis, a model that looks 3x cheaper on paper can end up costing only about 10% less once real token usage is counted.
  • Some smaller, more token-efficient models finish tasks at half the true cost of a 'cheaper' model, despite a higher advertised per-token rate.
04The hallucination benchmark
  • A hallucination-resistance benchmark that rewards honest uncertainty over confident guessing matters most for ambiguous, judgment-heavy work — raw coding benchmarks don't capture that risk.
05Head to Head: Kimi K3 builds
  • A model can win on functionality while still losing on visual and interaction polish — judge deliverables on the full experience, not just whether the feature technically works.
06Head to Head: Fable 5 builds
  • The most expensive model per-token can still be the best value once you weigh output quality against how much work it saves you.
07Head to Head: GPT-5.6 builds
  • A minimalist aesthetic doesn't excuse poor usability — small unreadable text and overlapping UI elements are real defects, not stylistic choices.
08Tallying tokens, time, and cost
  • A model can need six times more tokens than a competitor and still end up cheaper overall, if its per-token price is low enough to offset the inefficiency.
  • Despite using far more tokens, a rock-bottom per-token price can still produce the lowest total bill — inefficiency and cost aren't the same thing.
  • The cheapest total price doesn't guarantee the best result — factor output quality into any cost comparison, not just the invoice.
09The Verdict
  • A genuinely capable open-weight model's real advantage may only be 'cheaper than the priciest closed model,' not 'cheaper than everything on the market.'
  • Generation speed is a real, practical cost independent of dollars — a 90-minute wait for one build can be a dealbreaker for iterative work regardless of savings.
Glossary

Terms worth knowing.

Open weight model
A model whose trained parameters are published so anyone can inspect or run them, as opposed to a closed API-only model. It does not mean the model is small enough to run on a personal computer.
Cost per Intelligence Index Task
A benchmark from Artificial Analysis measuring the actual dollars spent to complete a standardized task, factoring in how many tokens a model needs rather than just its advertised per-token price.
Omniscience Index
A hallucination-resistance benchmark that rewards correct answers, penalizes confidently wrong ones, and scores an honest 'I don't know' as neutral, measuring whether a model overreaches on questions it can't answer.
Token efficiency
How many tokens a model needs to complete a given task. A model with a low per-token price can still produce an expensive bill if it is inefficient and needs far more tokens than a competitor.
Resources

Things they pointed at.

03:01toolArtificial Analysis ↗
01:43toolFrontend Code Arena (Arena.ai) ↗
05:38toolClaude Code
05:38toolCodex
13:26productChase AI community (Skool) ↗
Quotables

Lines you could clip.

02:54
“These open source Chinese models tend to be token hogs.”
single-sentence thesis of the whole video→ TikTok hook↗ Tweet quote
12:53
“It is so slow, which is kind of a deal breaker for a lot of people depending what you're doing.”
blunt closing verdict→ IG reel cold open↗ Tweet quote
12:42
“Kimi k three is an open source model that can compete with Fable, 5.6. However, it's not as cheap as you would expect.”
clean, nuanced summary line→ 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.

metaphoranalogy
00:00Did a Chinese open weight model just overthrow GPT 5.6 and Fable five? Well, if you've been anywhere near YouTube or Twitter in the last twenty four hours, then you have been bombarded with these charts showing Kimi k three performing just as well, if not better, than the best Anthropic and OpenAI has to offer, all while doing it at a fraction of the price.
00:20But should you believe the hype? Well, that's exactly what we're gonna answer in today's video as we take a deeper look at what these benchmarks actually tell us, and we do a head to head test between g p t 5.6, Fable five, and k three in the one area that Kimi k three is posting better numbers than anybody, front end design.
00:40So by the end of this video, we'll have a much better idea if Kimi k three is the new king. So let's begin with the numbers and benchmarks when it comes to Kimi k three because there's a lot we can actually pull out of these, and there's some nuance involved that goes beyond these charts that you've seen floating around all over the place, these two being the most popular.
00:57Now when we talk about Kimi k three, it is a 2,800,000,000,000 open weight model. When we talk about open weight, we talk about open source.
01:05This does not mean you can run this on your computer. We're talking millions of dollars of hardware required to run one of these things. But it's still great that it's open weight, that we can actually see how all the parameters are tuned versus something like Fable five or GPT 5.6 where that is invisible to us.
01:21Now how does it hold up on the numbers? This comes to us courtesy of Kimi and their moonshot lab, and it shows it competing, again, with Fable five and GPT 5.6 on all these major programming benchmarks.
01:34The other chart you've seen floating around is this one. This is coming from Arena dot ai, and it's measuring their front end code. Now how this benchmark works is if you go to arena.ai and you ask it, hey.
01:44Create me a landing page, it will give you two responses from two different models, and you don't know which one is which. It's like the Pepsi challenge, and you choose whichever one you want. Over time, we see which ones have been voted against its competitor.
01:57And right now, Kimi k three has done better than every other model in this space. But understand this is very, very subjective. And so when we look at these two charts and we compare that to the pricing between these models, we see that Kimi k three is also extremely cheap.
02:13Its input is $3 per million tokens compared to Fable, which is 10, compared to 5.6, which is 5. So it's 60% of the cost of 5.6, and it's 30% of the cost of Fable five on the input. And output wise, it's half the cost of 5.6, and it is 30% of the cost of Babel five.
02:32So way cheaper, and it's basically just as good if not better. So what's not to love?
02:37But this is where we need to start diving a little bit deeper in terms of cost and speed because it's one thing to see, oh, per million tokens, it's $3 and that one's 10. Well, how many tokens does it actually use? Not all models are created equal in terms of token efficiency.
02:49In fact, these open source Chinese models tend to be token hogs, which means while on paper, it's very cheap. What does that mean in reality?
02:57How does it compare to these guys who are actually very, very token efficient like GPT 5.6? Well, if we look at this from artificial analysis, which is an independent benchmarker, they have a cost per intelligence index task.
03:09So how much does it cost to complete these tasks? So Kimi k three right here, looking at 95¢ versus GPT 5.6 SOL is a dollar 4.
03:20So slightly more expensive. If we look at 5.5 extra high, you know, it's over here.
03:26If we look at Terra, which is 5.6 on max, it's half the price of Kimi k three. Now versus Claude, right, pretty expensive.
03:35Fable five, dollars 2.75. So this cost savings definitely applies when we compare Fable to Kymi K3. But when we compare it to 5.6, not a huge difference.
03:46We're talking like 10% or less. And in fact, there are certain instances where Terra, some of these smaller models in that 5.6 realm, are actually half the cost of the Kimi k three.
03:56Now the other thing we wanna take note of is time because time is certainly a currency we wanna pay attention to. Historically, these open source models from China are slow. You can see that here.
04:05The slowest model of the bunch is the old Qimi k 2.6. Now Qimi k three has gotten significantly better.
04:12In this chart, it's tracking at six minutes. If we compare that to Fable five, Fable five is at five minutes, and 5.6 is at 4.7. So, you know, it's less token efficient, and it's slower.
04:27But overall, much cheaper than Fable five and about on par with 5.6.
04:34Last benchmark I wanna talk about is the omniscience index. This is measuring hallucinations. They give the model a series of very difficult questions that it may or may not know the answer to, and it's graded in terms of did it get it right, did it get it wrong, or does it say I don't know?
04:49You get a point if you get it right. You lose a point if you get it wrong, and you get a zero if you just say I don't know. And we can see here, and this is out of a 100, Fable five scores the best at 40 points.
05:005.6 is way behind at 22, and then Kimi k three is at 18.
05:06And I think this benchmark is really important if you're using these agents in a context that just requires a lot of nuance and there's a lot of gray area. When we take all these benchmarks together in the aggregate, I think what we should walk away from is that Kimi k three on paper is extremely powerful. However, when we talk about how cheap it is, that's rather overblown, especially when we compare it to the GPT models that are hyper token efficient.
05:27On top of that, Kimi k three is just a tad bit slower as well. So then it just becomes a question of what of those three things is most important to you in regards to power, cost, and speed.
05:38So let's go to the front end testing between these three models. So right here on the left, I have Fable five running inside of Claude Code. And on the right, I have Kimi K3 also running inside of Claude Code.
05:49And back here, I have Codex. So this is the prompt I'm going to give it. I'm telling it I wanted to build a three d globe travel dashboard that feels like you've seen from my sci fi film, not just a website.
06:01I'm telling it I wanted to essentially make it a visual spectacle, and I wanted to get rather creative. In fact, I say surprise me with at least one idea I haven't seen before.
06:10I try not to get too specific with this prompt because I kinda wanna see how all these models diverge and what they actually build me. So let's put them to the test.
06:21Alright. So all three are done building their website. So we're gonna go through all of them, and then at the end, we'll compare the actual cost, the amount of tokens used, and the time.
06:30So let's start with Kimi three. So here we go.
06:35We have the three d Earth. I can zoom in and out. I can move it around.
06:41Uh, not super high fidelity, but that's alright. If I click on something like Mexico City, nothing really happens.
06:51What else? It has some other points on here, like Cape Town. Um, over here on the left, it shows some active routes.
06:58So if I click on an active route like Tokyo, it actually brings me to it. And then over here on the right hand side, I'll move over for a little bit, We can see, you know, it's lat long, time, weather, the best window to travel there, entry, and then sort of like how much it costs right now.
07:16And as I click through stuff over here on the left, you know, it brings me to these different towns as well as all their stats.
07:23So overall, pretty cool, and you can see how it also has the Earth kind of, like, divided into, you know, daytime versus nighttime and all that. So overall, it looks pretty sweet. Now let's take a look at Fable five.
07:35So right off the bat, I would say these graphics look a little bit cleaner. Right? If we compare these two, this just looks a little more crisp.
07:43Um, we also have some lighting, which is cool. We have that same, like, daytime nighttime breakdown, but I can actually I have a scroll wheel, a little thing right here where I can switch between day and night, which is pretty awesome.
07:57I can't zoom in as far, but I can see more of these sort of cities, and it looks a little bit more high fidelity. So if I click on one of these alright.
08:06Same thing as before. Brings up some stats. Lat long shows the conditions in terms of weather and time and then sort of what the fare is.
08:15Shows return fare, although I'm not totally sure, like, where this is actually coming from. If I click here over on the left, same sort of thing.
08:22Brings me to that city. I can see some of the stats as well as what it would cost. I think overall, in terms of sort of the, I don't UI polish between these two, I think I prefer what Fable five put out, especially this sort of, like, you know, day versus night thing it has happening.
08:44And interesting enough, it looks like when I change like where it is during the day, you can see it updates the cost on the right hand side depending on when you would actually be landing there.
08:55So pretty good. And now let's look at GPT 5.6. So initial thoughts, feels like a step down from what we saw with Kimmy k three and Fable in terms of the globe itself.
09:06Like, these globes definitely have a lot more going on and are much more of a visual spectacle. This, I think, they almost went more of, like, a sci fi look. It kinda has, like, random it's like, are these supposed to be okay.
09:20So same sort of thing. If I click on certain cities here, like Singapore, this pops up, but it kind of overlaps the globe globe itself, and everything is so small text wise, it's kind of hard to even read. I do have here at the bottom sort of like a day versus night kind of sweep like we had over there with Fable five, but it's not as great.
09:42And when it's nighttime, you, like, can't even see the continent itself. I guess those lights are supposed to be like cities. If I click over here on the left, yeah, same as before, brings us to those actual cities.
09:53Um, but, again, definitely feels like a step down from the last two we saw. So overall, when we compare these three, I would put Fable five in the lead. Not far behind, I would have Kimi k three.
10:04And then in third place, I would put GPT 5.6. So now let's compare tokens, time, and cost because we saw the end result, but what did we actually have to pay for it?
10:15So we had Kimi, we had Fable, and then we had Codex, which I'm just gonna put as x.
10:21So in terms of tokens, Kimi was by far the heaviest.
10:26It took it 21,500,000 tokens to come up with that.
10:31And in terms of time, it took one hour and thirty three minutes.
10:38So over ninety minutes to create that in 21,500,000 tokens. Cost, and this is all coming from Open Router, was $8.66.
10:47Now let's talk about Fable. In terms of tokens, it used 3,500,000.
10:53In terms of time, it took just seventeen minutes, and total cost was $11.64. So, again, what did I talk about at the beginning?
11:03We talked about speed and token efficiency. Kiwi k three is clearly way behind Fable in that regard, especially when it comes to time.
11:11And so when we looked at that token to token cost, like, per, you know, million, it was like, oh, it's a third of the cost of Fable. Well, in reality, not so much.
11:20Still cheaper than Fable, don't get me wrong, but not nearly as efficient. Then we had Codex, which I was kinda disappointed with its output, to be totally honest.
11:29It took 5,600,000 tokens. For time, it took twenty five minutes, and its total cost was $5.66.
11:38If you're wondering, like, tokens and cost and why doesn't that exactly line up for, like, inputs versus outputs versus what's advertised, understand there's a lot of caching going on for all three of these. So what does this kind of mean?
11:52Well, big thing is, Kimmy, right, very heavy on the tokens, and it took forever.
11:59Took an hour and a half. To be honest, when I was making this video, was like, oh, maybe I'll do, like, you know, three, maybe even four examples. No.
12:06I'm not sitting here for twenty hours to do this. And, you know, so way behind in those two places against the Frontier models, but overall, Talbot was solid. You know, I thought it did a really good job, all things considered.
12:18And the most expensive of the bunch, obviously, was Fable with Codex being the cheapest. So what can we really pull from all this?
12:27Well, I think what we can pull at least here on the front end side, and it probably applies the same to the coding if we wanna take the benchmarks at face value, and that's Kimi can compete. Kimi k three is an open source model that can compete with Fablem 5.6.
12:42However, it's not as cheap as you would expect.
12:46In certain cases, it's more expensive, and it's just slow. It is so slow, which is kind of a deal breaker for a lot of people depending what you're doing.
12:58But if that's okay with you, you know, then maybe it isn't so much of a downside.
13:04But I think I think the big thing that kinda gets lost in the benchmarks of the numbers is really the cost. Right? Because it's not as cheap as you would expect because of this inefficiency when it comes to actually using tokens.
13:15So that's where I'm gonna leave you guys. Hope I was able to shed some more light on Kimi k three. I think it's awesome we have an open source model that can compete, but let's not get ahead of ourselves in terms of its raw power and specifically its cost, both in terms of money and time.
13:29So let me know in the comments what you thought. Make sure to check out Chase A plus if you wanna get your hands on my Claude Code masterclass that also includes a Codex masterclass these days. And besides that, I'll see you around.
The Hook

The bait, then the rug-pull.

The title makes the claim outright: an open-weight Chinese model just beat Fable 5. The video spends its first minute stacking up the viral benchmark charts fueling that claim, then spends the other twelve minutes quietly taking it apart with a real build test and a real invoice.

CTA Breakdown

How they asked for the click.

VERBAL ASK
13:26product
“Make sure to check out Chase A plus if you wanna get your hands on my Claude Code masterclass that also includes a Codex masterclass these days.”

Single soft mention at the very end, after the full analysis is delivered — no mid-roll interruption, no repeated asks.

MENTIONED ON CAMERA
03:01toolArtificial Analysis ↗
01:43toolFrontend Code Arena (Arena.ai) ↗
13:26productChase AI community (Skool) ↗
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Storyboard

Visual structure at a glance.

cold open
hookcold open00:00
benchmark charts
valuebenchmark charts01:01
Kimi K3 build
valueKimi K3 build06:33
Fable 5 build
valueFable 5 build07:53
GPT-5.6 build
valueGPT-5.6 build09:13
cost tally
valuecost tally10:33
CTA / close
ctaCTA / close13:26
Frame Gallery

Visual moments.

cold open00:00Frame at 00:12 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:12Frame at 00:25 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:25Frame at 00:40 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:40Frame at 00:46 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:46benchmark charts01:01
cold open00:00Frame at 00:12 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:12Frame at 00:25 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:25Frame at 00:40 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:40Frame at 00:46 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King00:46benchmark charts01:01Frame at 01:12 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King01:12Frame at 01:22 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King01:22Frame at 01:34 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King01:34Frame at 01:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King01:43Frame at 01:53 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King01:53Frame at 02:04 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:04Frame at 02:14 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:14Frame at 02:25 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:25Frame at 02:33 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:33Frame at 02:45 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:45Frame at 02:56 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King02:56Frame at 03:06 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:06Frame at 03:17 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:17Frame at 03:27 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:27Frame at 03:38 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:38Frame at 03:48 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:48Frame at 03:56 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King03:56Frame at 04:09 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King04:09Frame at 04:19 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King04:19Frame at 04:29 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King04:29Frame at 04:40 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King04:40Frame at 04:51 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King04:51Frame at 05:01 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:01Frame at 05:11 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:11Frame at 05:22 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:22Frame at 05:32 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:32Frame at 05:38 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:38Frame at 05:53 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King05:53Frame at 06:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King06:03Frame at 06:13 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King06:13Frame at 06:23 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King06:23Kimi K3 build06:33Frame at 06:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King06:43Frame at 06:57 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King06:57Frame at 07:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King07:03Frame at 07:15 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King07:15Frame at 07:18 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King07:18Frame at 07:32 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King07:32Frame at 07:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King07:43Fable 5 build07:53Frame at 08:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:03Frame at 08:13 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:13Frame at 08:23 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:23Frame at 08:33 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:33Frame at 08:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:43Frame at 08:53 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King08:53Frame at 09:06 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King09:06GPT-5.6 build09:13Frame at 09:23 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King09:23Frame at 09:33 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King09:33Frame at 09:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King09:43Frame at 09:56 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King09:56Frame at 10:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King10:03Frame at 10:13 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King10:13Frame at 10:23 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King10:23cost tally10:33Frame at 10:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King10:43Frame at 10:53 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King10:53Frame at 11:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:03Frame at 11:13 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:13Frame at 11:23 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:23Frame at 11:33 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:33Frame at 11:43 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:43Frame at 11:53 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King11:53Frame at 12:03 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:03Frame at 12:15 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:15Frame at 12:24 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:24Frame at 12:34 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:34Frame at 12:44 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:44Frame at 12:54 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King12:54Frame at 13:04 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King13:04Frame at 13:15 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King13:15Frame at 13:24 from Kimi K3 vs Fable 5 vs GPT-5.6: The Real Cost of the 'Cheap' Open Source King13:24CTA / close13:34
Watch next

More from this channel + related breakdowns.

29:32
Theo - t3․gg · Essay

The weird situation with Fable

Theo breaks down how Anthropic silently modified prompts, rewrote its system card, and built invisible safeguards into its most capable model - then got caught.

June 15th
Chat about this