A single detailed prompt file run through an agent in a loop can autonomously research, script, art-direct, and render a full Vox-style animated explainer video using outside video, music, and voice models, for about $15 in generation costs.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You use an agentic coding tool like Claude Code and want to see a real, non-toy example of a multi-hour unattended run against one detailed prompt file.
You make explainer or motion-graphics content and want a grounded read on how close AI video models are to replacing an animator.
You're evaluating Replicate as a single API aggregator for stitching together several AI models, video, music, and voice, into one pipeline.
SKIP IF…
You want a beginner, click-by-click tutorial. This assumes familiarity with agentic coding tools, API keys, and .env files.
You're looking for a specific finished Vox-style video rather than the process used to build one.
TL;DR
The full version, fast.
Pat Simmons gave Claude's Fable 5.1 agent one prompt.md file and told it to loop until it matched a Vox explainer end to end. Fable pulled reference Vox videos with yt-dlp, sampled burst frames to study motion rather than just still frames, wrote itself a reusable style skill, picked a culturally relevant topic (record heat versus the 1936 Dust Bowl), wrote the script, then called Seedance 2.5 for video, Google Lyria 2 for music, and Minimax Speech 2.5 HD for voiceover, all through the Replicate API. After hitting its own usage limits and needing a credit top-up, it delivered a full Vox-style video for about $15, close but not indistinguishable from human motion-graphics work.
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.
Cold open plays the finished Vox-style video, then Pat explains that one prompt to Fable 5.1 produced it, replacing what used to take an illustrator, an animator, and tens of thousands of dollars at an ad agency.
01:00 – 01:40
02 · The prompt
Pat pastes the entire instruction into Fable as a single slash-goal command: read prompt.md end to end, don't stop until it hits feature parity with a Vox explainer.
01:40 – 03:33
03 · Study how Vox actually moves
Walkthrough of the prompt.md outline (goal, get the footage, how it moves, burst sampling, write the skill, pick the topic, render it, QA itself), then the footage and motion-study sections: extracting reference Vox videos via yt-dlp and burst-sampling frames instead of feeding full video so the agent can learn how things move, not just how they look.
03:33 – 03:54
04 · Making it write its own skill
After analyzing the reference videos, the prompt has Fable write a reusable skill file documenting the Vox visual style for itself and for future agent runs.
03:54 – 04:17
05 · Picking the topic and writing the script
The prompt tells Fable to pick a culturally relevant topic itself, verify facts against real sources, and write the script without checking back in.
04:17 – 05:26
06 · Setting up Replicate
Replicate is introduced as an API aggregator for AI models; Pat generates an API key and stores it in .env so the agent has direct access to call Seedance 2.5 and the other models it needs.
05:26 – 06:02
07 · Music and voiceover
The pipeline adds Google's Lyria 2 for a subtle background music bed and Minimax Speech 2.5 HD for the narrated voiceover, called through the same Replicate API.
06:02 – 06:14
08 · QA rules
The final prompt section tells Fable to burst-sample and critique its own output, looping until the result actually looks right before sharing it back.
06:14 – 06:57
09 · Fable starts building
Checking in hours later: Fable has extracted the Vox style, built a reference contact sheet, researched a topic, and locked the script around record summer heat versus the 1936 Dust Bowl, and is now generating keyframes.
06:57 – 07:49
10 · Hitting usage limits
Fable maxes out its usage limits mid-run. Pat has to buy additional credits to let it finish, a real cost the demo doesn't hide.
07:49 – 08:49
11 · The reveal
The finished video plays in full: the July 1936 heat record, the Dust Bowl, this year's broken record, and the closing line that this one has no fix that simple.
08:49 – 10:50
12 · Where the AI tells show
Pat rewatches the video scene by scene, calling out a slightly off frame rate, transitions that land well, a shaky stop-motion glitch, and a version of Vox's posterized-time effect that doesn't fully land, alongside praise for the topic choice and pacing.
10:50 – 11:48
13 · What Fable actually did
Recap of the full run: pulled and analyzed Vox reference footage frame-burst by frame-burst, wrote a style skill, chose and researched the topic, wrote the script, then rendered scenes through Seedance 2.5 via Replicate and layered in Lyria 2 music and Minimax voiceover.
11:48 – 12:13
14 · Swapping the models
The two easiest things to change: the video model itself (Seedance 2.5, Gemini Omni, Runway, Kling) and the target visual style (Vox, whiteboard, retro cutout), without touching the rest of the prompt.
12:13 – 12:28
15 · What it cost
Full cost breakdown: about $14 for the one-minute Seedance 2.5 render at 720p, with music and voiceover included, for a total of roughly $15.87 to generate the finished video.
12:28 – 13:08
16 · Where this stops
Pat's honest caveat: a trained eye can still see the tells, and this doesn't replace a real motion designer at production scale. Where it's genuinely useful is a marketer, designer, or solo creator bringing an idea to life without a full production budget.
13:08 – 13:44
17 · Outro
Pat points to a blog post with the full prompt.md, style guide, and finished video, notes Fable 5.1 did noticeably better than Fable 5 on the same kind of task despite being cheaper, and signs off.
Atomic Insights
Lines worth screenshotting.
A single prompt.md file can drive an agent through research, scripting, art direction, and video rendering in one autonomous loop, with zero back-and-forth questions allowed.
Feeding a video model full video files for style analysis doesn't work, it just extracts frames one by one and misses how things move; sampling bursts of frames every few seconds captures motion instead of just look.
Having the agent write its own reusable style skill after analyzing reference videos means future videos in that style skip re-doing the analysis from scratch.
Replicate works as a single API aggregator for video (Seedance 2.5), music (Google Lyria 2), and voice (Minimax Speech 2.5 HD) models, so one agent can call all three without juggling separate accounts.
A one-minute 720p Seedance 2.5 generation cost about $14; the full video with music and voiceover came to roughly $15 total.
The agent hit its own usage limits mid-run and needed a manual credit top-up to finish, a real cost most AI-agent demos leave out.
Even a strong one-shot AI result still has visible tells: the frame rate sits a little off, Vox's signature posterized-time effect isn't fully reproducible yet, and there's a hint of leftover After-Effects-style structure underneath the animation.
The two easiest levers to swap without touching the rest of the pipeline are the video model (Seedance 2.5, Gemini Omni, Runway, Kling) and the visual style (Vox, whiteboard, retro cutout).
What used to take an illustrator, an animator, and tens of thousands of dollars at an ad agency can now come out of one prompt for about $15.
The real bottleneck isn't the video model anymore, it's writing a prompt detailed enough that the agent never needs to come back and ask a question.
Fable 5.1 handled this materially better than Fable 5 did on the same kind of task, mainly through tighter QA looping, even though it's the cheaper model.
Takeaway
One detailed prompt file can run a full video production loop by itself.
WHAT TO LEARN
A single prompt.md, written once and handed to an agent with a strict no-questions-asked rule, can chain research, script writing, art direction, and multi-model rendering into one unattended run that costs about $15.
01Intro
What used to take an illustrator, an animator, and tens of thousands of dollars at an ad agency can now come out of one prompt for about $15.
02The prompt
The entire build instruction is one slash-goal command pointing at a single prompt.md file, with an explicit don't-stop-until-parity rule.
03Study how Vox actually moves
Feeding a model full video files for style analysis doesn't work; burst-sampling frames every few seconds captures motion instead of just look.
04Making it write its own skill
Having the agent write a reusable skill file after its style analysis means future runs in that style skip re-doing the research.
05Picking the topic and writing the script
The prompt explicitly tells the agent to choose its own topic and verify facts against real sources rather than checking back in.
06Setting up Replicate
Replicate acts as a single API aggregator, so an agent can call multiple providers' models with one account and one key.
07Music and voiceover
The same pipeline layers in a dedicated music-generation model and a separate text-to-speech model for voiceover, both through Replicate.
08QA rules
Telling the agent to burst-sample and critique its own output before sharing it back is what turns a rough first pass into a finished result.
09Fable starts building
Full autonomous runs like this take hours, not minutes, even once the prompt and API access are fully set up.
10Hitting usage limits
Budget for the agent's own usage limits, not just the external API costs; a real run may need a mid-run credit top-up.
11The reveal
A genuinely well-produced result is possible from a single prompt with no manual storyboarding or scene-by-scene direction.
12Where the AI tells show
Even strong output has visible tells: an off frame rate, an imperfect signature effect, and traces of underlying structure that don't quite match hand-built animation.
13What Fable actually did
The full chain, footage analysis, skill writing, topic research, scripting, and multi-model rendering, ran as one continuous loop with no handoffs back to the human.
14Swapping the models
The video model and the target visual style are the two cleanest levers to change without touching the rest of the prompt.
15What it cost
A one-minute 720p video render cost about $14; music and voiceover added roughly $1-2 more, for about $15 total.
16Where this stops
This replaces the cost of a production, not the eye of a human motion designer; it's most useful for ideas, marketing, and presentations, not client-facing production at scale.
Glossary
Terms worth knowing.
Fable 5.1
An autonomous coding agent, driven here with a slash-command goal, that can run for hours against one prompt file, researching, writing, and calling outside APIs without further input.
Seedance 2.5
A text-to-video AI model, called through Replicate, used here to render the individual Vox-style animation clips.
Replicate
An API aggregator that hosts many different AI models, video, image, music, and speech, behind one platform, so an agent can call several providers with one account and one API key.
Google Lyria 2
A text-to-music generation model used here to create the subtle background music bed for the explainer video.
Minimax Speech 2.5 HD
A text-to-speech model used to generate the narrated voiceover for the finished video.
Burst sampling
Pulling a rapid sequence of frames every few seconds, instead of full video, so an AI model can study how something moves rather than just what a single frame looks like.
Vox style
The visual language of Vox's animated explainer videos: paper-cutout collage animation, torn-paper title cards, and a posterized-time motion effect, which the agent was set to study and imitate.
“When I worked in ad agencies, a video like this took an illustrator, it took an animator, it took hundreds of hours inside of After Effects, and we would pay tens of thousands of dollars for something just like this. Today, all it comes down to is a single prompt.”
states the whole video's before/after thesis in one line→ TikTok hook↗ Tweet quote
07:20
“Anthropic's got a gun to my head and I really have no choice but to buy credits.”
funny, honest admission of the real cost behind an 'autonomous' agent demo→ IG reel cold open↗ Tweet quote
08:56
“It definitely feels less natural than Vox.”
candid, one-line verdict that undercuts the hype without dismissing the result→ newsletter pull-quote↗ Tweet quote
12:25
“Total is about 15 bucks to generate this one minute video.”
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.
17px
metaphoranalogystory
For 90 years, the hottest month ever measured in America was July, 1936. The Dust Bowl. That July, North Dakota hit 121 degrees.
The plains had been plowed to bare dirt, and bare dirt can't cool itself the way grass can. When I worked in ad agencies, a video like this took an illustrator, it took an animator, it took hundreds of hours inside of After Effects, and we would pay tens of thousands of dollars for something just like this. Today, all it comes down to is a single prompt.
I didn't storyboard anything. I didn't write prompts for every scene. I didn't even know how to describe this Vox style.
All I did was take this to Fable 5 .1 and have it run with basically everything. So in this video, I'm going to show you exactly how to do this yourself and create Vox style motion graphics with Fable 5 .1 and Seedance 2 .5. So not wasting any more time, let's get right into it.
I'm gonna paste in this prompt and I'm gonna show you exactly how we build this ourselves. And here's the prompt we're giving to Fable. You ready for how complex this is about to be?
All we're saying is slash goal. read prompt .md followed end end don't stop until you hit feature parity with a vox explainer everything else lives in the file so we're sending fable on a loop here with slash goal and we're laying out all the details we have a very detailed prompt in that prompt .md and here's what that prompt .md looks like i'm going to break this down into some separate slides here because it is very extensive but we're going to go through each one of these sections so i can call this out to you but if you don't want to listen to me ramble you can also just copy this and give it to your own agent but i'm going to break down exactly what we're doing here because it is super important to have have a nice detailed prompt for the agent to adhere to especially when you're doing something as complex as creating box style motion graphics so first up in this prompt .md we have the goal I'm just restating this again not totally necessary but I'm just saying you're doing all of this the research the analysis the writing the art direction the actual API calls to the video model nothing in here gets handed back to me just to emphasize don't come back to me and ask any questions just get it done and then we have making a two minute narrated explainer that a person could mistake for a Vox video not Vox inspired emphasizing feature parody and so
it is going to extract in this next section here. We'd say, here's how to get the footage. We're using what's called YTDLP, which is extracting from YouTube videos.
You need to actually use the cookies in my browser. Otherwise you'll get these 403s. So I just call this out, not getting too technical.
Again, you can just feed this into your own agent and we'll just extract this properly if you tell it to do it just like this. And what we're doing is we're extracting like four to five Vox videos. So Fable can thoroughly analyze this and get a sense of, okay, this is what Vox talks about.
This is how they write their script. scripts. And also this is how they do their animations.
So that's actually extracting the footage, doing the research. And then I mentioned here, study how it moves, not just how it looks. The one call out with a model like Fable is you can't feed it full videos.
If you do, it's just going to extract it via something like FFmpeg and it's going to analyze each screenshot one by one. So what I'm doing here, because I know that won't really work, you won't get the essence of a Vox style explainer without saying to sample a frame every few seconds instead. So it's actually taking a burst of frame so we can kind of analyze exactly what that movement is.
I just call out some things like, you know, what eases, how hard it overshoots, stuff like this that you would only notice when you're watching a video, camera movement, all of that kind of stuff, even like how long a beat holds in a certain animation. So that's how it moves. And then the next is just sampling.
Like I mentioned, FFmpeg, we're just getting burst, stitching the bursts into a context sheet to read it, to get a sense of how these move. And then after we've done a thorough analysis of Vox's style, I'm asking it to then create some sort of template.
I'm just saying, write a skill so that another agent could follow this. I want to use this in the future as well. So after we've built out and defined this visual style, we'll have a nice specific skill for an agent to reference in the future.
So that's what I'm asking to next, creating a skill for VoxStyle explainers. And so once we have that skill written, we're ready to start constructing this with the video model. And we'll just start from the beginning.
So we'll start from the concept, which is in this case, I'm just telling it, pick the topic. I want it to be like Vox too, in that it's culturally relevant, something that is interesting, just like Vox does. I want it to just pick it, and I want it to look up the sources and just research this a bit.
And then once it's researched, once the script is written, once the beats are laid out, then it's ready to go to rendering this. And to actually call this API, we're going to be using a service called Replicate. Replicate is, if you just go to replicate .com, is an aggregator of APIs.
And it's super nice, so you don't need to go to all of these different companies and get APIs. there you can just test all of these in one platform and use it directly through replicate so there's a number of these aggregators there's foul i mean technically you could probably use higgs field too i just prefer replicate because i've used it before and so i have an account and everything we're going to go ahead and just sign in with github and then once you sign in you'll have this dashboard here and you can literally just generate the api i don't even like to go searching for these models or anything i just give it to claude and have it run with it you can see i've loaded up some credits so make sure you do the same and then just get your api key click your profile on the top left go to api tokens and then just create a new token copy that token store it in dot env and your project files and just like that your agent has full access to replicate and then once you have the api with your agent it's just going to call the cdns 2 .5 so you can actually just search here too just to show you c .5 show it's just going to call this api here replicate makes this very easy for an agent to read too so the agent will know exactly how to do this to actually access this video model this golden retriever puppy
Look how real that is. That's crazy. And we're also doing some voiceover.
So in this case, we're using Google's model. It's called Lyra 2. Here it is.
Lyra 2 music generation model. So it's just a simple music generation model. We're just going to have a music bed, just very subtle in the background of this Vox video, exactly like Vox would do.
And then also, of course, we need voiceover. So we are using Minimax's model for this. Minimax speech 2 .8 HD.
You can use any number of text of voice models. At this point, they're so good. You can use your 11 labs API for you.
11 Labs and really any of these I just like using minimax because it's cheap I've tested this before and it's pretty good so that's replicate that's everything that's going to be generated with the replicate API and finally we just have QA so I'm asking fable to do the exact same thing just burst sample your own output and just be very strict about QA and really just go in a loop until this is done it looks good and then share it with me all right checking in on fables progress it has been humming for a few hours now so much so that I went to the gym and that's why I have a different shirt on but it looks like it Extracted the Vox style here and went through a few different videos pulled some reference clips put together a contact sheet did some more research Across the board on a potential topic and it found a potential topic somewhere.
Yeah actual research in the script topic is locked It looks like it's gonna talk about some sort of global warming thing. It looks like at 120 degrees Yeah, so that's locked and then it's just going through and it's generating these keyframes. So we'll give it some more time It's gonna take white a while to actually generate these and then qa this and probably revise based on feedback and just continue to go in a loop all right well we just hit our usage limits no surprise with fable 5 .1 just absolutely maxing out usage limits and not it's not like it was that much work i mean this is this is the downside of working with fable it's like geez it just eats through your usage however i think it's almost done it says ship ship as is it looks like it found some things i'm just not even going to worry about these like little things these two seams stay soft i don't even know what it means by that so i'm just going to see ship as is and let's see if it will actually generate this okay i'm out of usage credits all right anthropics got a gun to my head and i really have no choice but to buy credits because i have another anthropic subscription that's also maxed out so hopefully this won't cost too much money but i'm gonna buy credits i think it's almost done so i should be able with just i don't know a few dollars wrap this thing up it's a good time too to say please subscribe to the channel maybe even like this video drop a comment help with the algorithm because uh anthropic is
earning a hole in my bank account continue okay that only costs a couple dollars so we have this full video now output available here let's open this up here it is in all its glory all right big reveal let's see how well fable did for 90 years the hottest month ever measured in america was july 1936 the dust bowl that july north dakota hit 121 degrees the planes had been plowed to bear dirt and bare dirt can't cool itself the way grass can.
This summer, that record fell. This July was hotter, by a tenth of a degree. And it wasn't one region.
Every state in the lower 48 ran hotter than normal, all at once. But the days weren't what broke the record. The nights were.
Overnight lows averaged 64 degrees, the warmest nights ever recorded, and no chance to cool off. 1936 was partly man -made. and it ended when the grass grew back.
This one has no fix that simple.
some AI dead giveaways. I mean, even the it's like this weird little shaky glitch. It's trying to do that sort of stop motion collage animation that Vox does really well.
But I mean, some of these transitions are killer. We have this like this paper here, a little crumble, that little shake. We have a map, of course, is very Vox.
I like this little icon, even though it's just a kind of like the way it's doing it. You know, it's like we're going to get nitpicky. It definitely feels less natural than Vox.
But I mean, even the background here, the grid, the like communicating the idea, the script was pretty good too. Like it chose a good topic.
I didn't know that July was the hottest month, I guess, on record. And then it's talking about the Dust Bowl. Like it's interesting too.
I didn't realize that the Dust Bowl had to get too in the weeds here, but the Dust Bowl, the reason why it got so hot was because there's no grass. Whereas we now have grass, we're still getting hotter than the Dust Bowl. And then this nice little, nice little fall here.
The bar chart's a little weird. It's trying to do, you know, that's a classic Vox thing, but it's a little glitchy right there. A map of course pretty good visual i like the little sharpie and then this is a nice touch too feels very collagey where it just kind of goes up that sun nice transition there and then it goes to night a little kind of weird whatever this is a little curtain drop here but this is nice too i mean the house comes in it's at night says the night drop to only 64 degrees on average we have this movement kind of just kind of weird you know if i just play this like i said a little glitchy it's like kind of staticky it doesn't feel as natural but i know what it was trying to do is trying to get across that kind of box movement and that's sort of i think it's called posterized time then boom and then reveal just really good i mean that was way better than i even thought it was going to be and that was a one shot like let me just show you this again this is what we did slash goal it went through extracted all this it got that vox style it created a style guide which i will give to you in the link below as well the style guide that created in analyzing vox videos went through rendered all the frames you have all these and the actual video here
Pretty impressive. Okay, so that is basically it. Let's wrap this thing up by just going through exactly what we just did.
So this is what happened. We had that prompt .md, which again will be available to you in the link below. And Fable just followed all of this.
So it pulled the footage, downloaded a handful of Vox videos, took them apart frame by frame. It read the motion, not just the static images it was extracting, but actually looking at. the frames stitched together in this burst frame so we kind of see how the movement played out it wrote a skill after analyzing these vox videos for itself and for any future agents to reference it thought through more about the topic and actually writing the script and kind of being on brand with vox it chose the topic then it started to after writing the script actually plan out the scenes So art directing the shots and then rendering them together by going to Cdance 2 .5 via the Replicate API and just generating them that way, then actually cutting it all together.
So generating the voiceover, little music bed with Google Lyria and Minimax. That's essentially it. So just a couple of caveats.
You don't need to stick to Cdance. You can just use whatever video model is the most popular. This is the beauty of Replicate.
You can test all of these, see which one has the best output, and then just run with that model. You could use Gemini Omni. You could use Runway.
You could use Kling. These are pretty good at this now. And then of course, you don't need to just do a Vox style.
You know, you can choose whiteboard or retro cutout, any number of different styles, depending on what type of video you want to make. And it really is just a prompt that Fable 5 .1 is falling to a T. Next, let's talk about costs.
And it was surprisingly cheap for this. For a one minute generation with Seedense 2 .5 at 720p, not 1080p, was about $14. And that was really cheap with the actual music and the voiceover.
So total is about 15 bucks to generate this one minute video. And then finally, just a quick caveat. I just wanted to call this out.
If you're... motion graphic artists watching this you can definitely see the tells and i know in the beginning i said we would pay you know tens of thousand dollars for this at ad agencies which this is pretty darn close but it still feels like it hasn't replaced humans and i don't want it to honestly i think the purpose of this and where it's fun is just being able to create these yourself to have an idea to kind of bring it to life in a different type of style and to maybe include it in if you're a marketer or if you're a designer or even if i don't know you're trying to spice up a presentation or something at work like that's what these video models could be used for like In true production, unless you're a small company, unless you're getting scrappy with it and just can't afford a motion designer, I still think that nothing beats the real human thing.
So that's going to do it. Like I said, everything is in the link below. There's a whole blog post with the prompt or the style guide with the full video if you want to watch that again.
But yeah, I'm curious. Start putting this to the test yourself. I would love to see some generations that people are doing with CDNs 2 .5 and all these new video models because it's really fun time to be creating this stuff when you have none of these skills, especially with Fable 5 .1.
I think it did particularly well. Even Fable 5. I've tried this.
before with Fable 5. It didn't do as good. 5 .1 is cheaper, even though I did hit my usage limits.
It's just a little bit smarter, and I think it did a much better job of QA and going in that loop and everything. That Fable 5 may have fallen slightly short. Anyway, that's gonna do it for this one, and I'll see you in the next one.
The Hook
The bait, then the rug-pull.
The video opens cold with the finished AI-generated Vox explainer itself, a clip about record heat and the 1936 Dust Bowl, before cutting to Pat Simmons explaining that a single prompt, not an illustrator or an animator, made it.
Frameworks
Named ideas worth stealing.
01:40list
The prompt.md structure
The Goal
Get the Footage
How It Moves
Burst Sampling
Write the Skill
Pick the Topic
Render It
QA Itself
The eight sections of the single prompt.md file Fable was told to follow end to end without stopping to ask questions, from stating the goal through a final self-QA loop.
Steal forany 'build me a video in style X, autonomously' agent prompt; swap in different reference videos and a different target style
CTA Breakdown
How they asked for the click.
VERBAL ASK
13:08link
“everything is in the link below, there's a whole blog post with the prompt, the style guide, and the full video”
Soft CTA at the very end pointing to a blog post with the reusable prompt.md, style guide, and finished video; also a mid-video aside around the usage-limit moment asking viewers to subscribe, like, and comment while joking about the credit spend.
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.
A blind, five-round test pits Opus 5 against Fable 5 and Opus 4.8 across web design, 3D, games, motion graphics, and a SpaceX investment deck — model names stay hidden until the ranking is locked in.
Fifteen parallel builds, three models, and a blind reveal system, testing whether Fable 5.1's benchmark gains actually show up in real creative and coding work.
A creator walks through cloning a $2 billion venture-backed dictation app in one Claude Code session, then benchmarks his free local build against the real thing.
Ten identical builds, five models, blind-ranked before the reveal — a real-world stress test of Moonshot AI's new open-source model against GPT-5.6 Sol, Opus 4.8, GLM 5.2, and its own predecessor.
A blind, four-way bake-off — GPT-5.6 Sol against Fable, Opus 4.8, and GPT-5.5 — across ten builds and knowledge-work tasks, scored one task at a time without knowing which model made what.
00:15
00:25
00:34
00:46
00:56
01:17
01:27
01:48
02:02
02:08
02:19
02:29
02:42
02:49
02:57
03:10
03:20
03:41
03:51
04:12
04:32
04:40
04:53
05:03
05:14
05:24
05:45
05:55
06:26
06:36
06:46
07:07
07:17
07:28
07:38
07:48
07:55
08:09
08:29
08:40
09:00
09:11
09:21
09:31
09:41
09:52
10:03
10:12
10:27
10:28
10:43
11:04
11:14
11:24
11:35
11:48
12:06
12:37
12:47
12:57
13:18
13:28
13:38
27:44
29:59
15:29
36:06
40:03
26:26