Modern Creator
Nate Herk | AI Automation · YouTube

I Tested Astra 6.2 vs. Fable 5.2 Across 12 Real Tasks

Twelve identical prompts, two models on high effort. Fable won seven on output quality. Astra won on cost, speed, and token efficiency.

Posted
3 days ago
Duration
Format
Review
educational
Views
113.4K
1.2K likes
Part of the collectionThe Claude Fable 5.2 PlaybookEvery Fable 5, 5.1 and 5.2 breakdown, synthesized into one page.
Read the playbook
Part of the collectionThe GPT-6 Astra PlaybookEvery GPT-6 Astra breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

At identical per-token prices, Astra 6.2 in Codex is far more token efficient than Fable 5.2 in Claude Code, so Fable's 7-to-5 quality edge costs roughly seven times more per task.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You pay for both Claude Code and Codex and want to know which model to point at which kind of job, backed by cost and runtime for each run.
  • You build websites, decks, videos, or dashboards with AI agents and want to see real outputs side by side before picking a default model.
  • You've built custom skills for your agents and want to know whether a new model will actually pick them up.
SKIP IF…
  • You want hardcore refactoring or large-codebase benchmarks. The author says none of these 12 tests were heavy software engineering.
  • You want a rigorous scored benchmark. The winners are the host's subjective calls on 12 hand-picked tasks.
TL;DR

The full version, fast.

Nate Herk gave Fable 5.2 (Claude Code) and GPT-6.2 Astra (Codex) the same /goal prompts on high effort across 12 tasks, including web design, slide decks, video edits, a 3D summit, trip planning, a financial model, and a browser game. Fable won seven on output and Astra won five, but Fable cost about 7.2x more in API terms and ran almost 5x longer even though both are priced at $8 input and $40 output per million tokens. Astra is better at vision and aesthetics and was cleaner on simple tasks. Fable is stronger on logic, detail, and following custom skills. His advice is to use Fable when quality matters most and to build skills so Astra can do the repeatable work cheaply.

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 – 00:48

01 · The 12 tests and setup

Same /goal prompts, separate work trees, high effort, and cost, tokens, and runtime tracked for each run.

00:48 – 03:18

02 · Web design

Two Perk Form sites. Astra's clean one wins at $2.50 vs Fable's $17.50.

03:18 – 05:59

03 · Slide deck

12-slide Hercules Advisor deck. Astra simpler, Fable busier; the cheaper one wins on a simple task.

05:59 – 07:56

04 · Canva painting

Astra's painting is far more realistic and it worked around a stroke limit. Fable $63 vs Astra about $7.50.

07:56 – 09:59

05 · Event recap video

Both 60-second recaps are bad. Fable is the lesser evil.

09:59 – 13:57

06 · 3D virtual summit

Fable's first-person world functions better and feels more on brand; Astra's looks better.

13:57 – 15:06

07 · Instagram Reel

Both good. Fable wins on sound design and B-roll at about four times the cost.

15:06 – 17:39

08 · YouTube edit

Astra ignores the host's skill and invents a cheap style. Fable uses the right skill but runs over five hours.

17:39 – 20:12

09 · Learning guide

Interactive guides from a tutorial video. Astra's is cleaner and about $3 vs $17.

20:12 – 22:34

10 · Trip planning

Astra looks better, Fable plans better with links and a working map. Fable wins.

22:34 – 24:15

11 · Instagram carousel

Neither model found the brand skill. Astra's is less overwhelming and cheaper.

24:15 – 26:21

12 · Financial model

Fable's workbook has more detail and logic; it costs $15 vs $2.50.

26:21 – 27:47

13 · Browser game

Fable's 3D game is actually fun; Astra's is slow and 2D.

27:47 – 29:38

14 · Final score and costs

Fable 7, Astra 5. Fable 7.2x cost, 4.7x runtime. Same per-token price.

Atomic Insights

Lines worth screenshotting.

  • Fable 5.2 and Astra 6.2 are both priced at $8 per million input tokens and $40 per million output tokens, so the cost gap comes entirely from how many tokens each agent burns.
  • Across all 12 tasks Fable cost about 7.2x more in API billing, ran about 4.7x longer, and used about 9.4x the input tokens.
  • Fable won 7 of 12 tasks on output quality, yet it was the more expensive choice on every single one.
  • Astra's aesthetic and vision work was consistently stronger: its Canva painting looked far more like the real photo than Fable's.
  • Fable's output tended to be better on logic and function, such as trip itineraries with links and financial models with driver tabs.
  • When Codex hit Canva's stroke limit it exported, re-imported, and kept painting. Fable hit the same limit and just stopped.
  • Astra ignored the host's custom YouTube editing skill and invented its own cheap style, while Fable used the right skill and produced the better edit.
  • Both models failed to find the host's Instagram carousel brand skill, so a skill that works on one model is not guaranteed to work on another.
  • For simple jobs like a 12-slide deck, the cheaper model wins by default because both outputs are good enough.
  • A busy slide deck suits an emailed read, and a simple one suits a projector. Judge outputs against how they'll be used.
  • The Fable YouTube edit ran over five hours of active work, which is far longer than the host's usual Opus 5.5 workflow.
  • Vague prompts make models show their defaults, which is why the host keeps his test prompts loose on purpose.
  • Both models independently planned the same trip route and named their browser games the same, which shows how similar their creative defaults are.
  • The host's practical strategy is to pay for Fable once to find a good output, then encode it as a skill and hand repeat runs to the cheaper model.
Takeaway

Pick the model by task, not by headline score

WHAT TO LEARN

At the same token price, the cheaper agent finished 12 tasks for about a seventh of the cost, so quality edges only pay off where the output really matters.

  • Equal per-token pricing does not mean equal cost: Fable used about 9.4x the input tokens, which is why its total bill was roughly 7.2x higher.
  • For simple jobs like a clean slide deck or a basic site, the cheaper model's output is good enough and the savings decide the winner.
  • Astra leans toward visual polish and strong vision work, such as a realistic Canva painting and cleaner learning guides.
  • Fable leans toward logic and detail, such as itineraries with links, financial workbooks with driver tabs, and a game that is actually fun.
  • Agents can silently ignore your custom skills, so test each new model against your skills before trusting it with a recurring job.
  • Judge an output against its use: a dense deck works as a leave-behind and a sparse one works on a projector.
  • A strong workflow is to pay for the best model once, capture what worked as a skill, then hand repeat runs to the cheaper model.
  • Own-use tests beat public benchmarks for deciding, but remember these 12 wins are one person's subjective calls, not hardcore engineering results.
Glossary

Terms worth knowing.

Astra
OpenAI's GPT-6.2 model as run through Codex. In this video it is the cheaper, more token-efficient contender.
Fable 5.2
Anthropic's model as run through Claude Code. In this video it wins more tasks but uses far more tokens.
/goal prompt
A task prompt given to an agent in a slash-goal style so it works autonomously toward an outcome. The host sends identical ones to both models.
Work tree
A separate checked-out copy of a project folder. Each model ran in its own so they couldn't interfere with each other.
Token efficiency
How few tokens a model spends to finish the same task. With equal per-token prices, higher efficiency means a lower bill.
Skill
A saved set of instructions and context an agent can load for a recurring job, like a brand style guide for carousels.
Resources

Things they pointed at.

00:00toolClaude Code
00:00toolCodex
06:00toolCanva
Quotables

Lines you could clip.

28:08
“Both of these models cost the same via API billing. It's just that Codex with Astra 6.2 is significantly more token efficient.”
Clear, surprising cost takeaway→ TikTok hook↗ Tweet quote
23:04
“Models are different and you have to make sure that your skills work with all your models because they always interpret everything a little bit differently.”
Practical lesson with no setup needed→ newsletter pull-quote↗ Tweet quote
25:22
“Astra has a really good feel for the experience aesthetically and Fable has a better feel for the logic and the actual journey.”
Compact model-personality contrast→ 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.

So I just put GPT 6 .2 Astra head to head against Fable 5 .2 across 12 real use cases. I'm talking things like web design, building professional slide decks, travel planning, 3D games, financial dashboards and spreadsheets, editing Instagram videos, editing YouTube videos, and a few other things as well. and at the end of each one of these tests we're going to go over how much it cost me in dollars input and output tokens as well as how long it took to run so that you actually know which model is best for what use case so i don't want to waste any time let's just get straight into this one and so the way that this is going to work is that i send them the exact same slash goal prompts so that they're in the exact same conditions they're on separate work trees they can't interfere with each other here is the prompt that i gave astra and you can see that astra and Fable are both going to be here running on high.
So they're both on high effort, they both get the exact same prompt, and now let's take a look at these sites. Okay, so experiment number one was web design. And we did perk form once again, which is just this fake brand that I have that I like to use because it's already got some brand guidelines and assets and things like that inside.
So I'm not gonna tell you yet which website was built by which model. You'll have to guess for yourself. This is website number one.
We can see right here we have everything on brand. We have a real photo right here. Coffee that pulls its weight.
We can see that we've got this little bit of a dynamic nav bar as I scroll up and down. We have one less thing to carry. And then this is a bit of a dynamic element here.
Although you can see when we go to the third image, we're kind of already like scrolling away. So that timing I think should be fixed a little bit, but not too bad. Now down here, there's an issue where when we're choosing between flavors, you can see that the image is kind of cropped out.
So it should have really used vision to know that we need to make this image slot a little bit larger. And then that's pretty much the end of the site. It's super clean.
There's not too much going on. A few minor issues, but it's not bad at all. This to me does not scream AI generated.
And here is site number two. I like that animation with the text that comes in. There's also like this little ring around this can that is orbiting.
We have a little bit of a layered effect here with the text and the can. As we scroll down, the can moves with us and we have your morning should not need two drinks. You can also switch flavors all the way up here.
And in the hero section as well, we can switch flavors too, which is pretty clean. We've got this interesting effect where we have coffee and protein and then they sort of swirl together, which is basically the whole value prop of... perk form and then we have a very dynamic site.
So it's not too much going on but it's definitely more movement than the previous site. You can see that these images do fit in the crop like they should. And then finally we just have the end of the site.
So place your bets before I reveal that this site was Fable 5 .2 and this one was Astra 6 .2. And here we have our stats. So Fable ran for 56 minutes while Astra ran for 12.
The input tokens were very, very different. The output tokens were very, very different. And if this was API billing, this would have cost me $17 .50 with Fable and only about $2 .50 with Astra.
So Astra here was significantly cheaper. So I want to say that I think that both of these sites are both very good. But ultimately, I think that in this case, I'm going to go with Astra 6 .2.
It's just super, super clean. Yes, there are a few issues like this and like this, but we could easily fix that, and I just think that this site is super clean, whereas this one does make you scroll through quite a bit, even though it is very well done. And of course, when you think about the fact that Astra was also this much cheaper, I want to give this one to Astra.
So right now we are 1 -0, Astra is in the lead. All right, so here is experiment number two. We built a slide deck.
So I said, hey, turn this mock survey into a 12 -slide Hercules Advisor presentation. So that's just another fake brand that I've got here that already has brand guidelines and things like that in there. So this is the prompt.
Let's pull up Astra's output first, and then we'll look at Fables. All right, here is Astra's report. We can see the coloring.
We can see the logo all looks good. The first live AI workflow. We have a nice chart.
We have not too much text going on, which is nice. It feels... Pretty professional.
It feels a little bit simple, but that's not a bad thing, right? More stats. Everything is on brand 100%.
We have more charts. It starts to feel a little bit of a repetitive format here, like three slides in a row, basically the exact same. Anyways, we've got more stats.
We've got more text down here. We've got columns. We've got different types of slides.
And then we've got number 11 and number 12 to close us off. So that is what this one looks like. Now here we have Fables output.
A little bit more going on. There's also like a chart. on the title slide, which I don't know if I love, but as we continue to go down, we have the branding in the top right of every slide, it looks like.
We've got a different sort of format here. There's definitely more words going on here, too. Some charts, some more hero text.
We've got different boxes and different columns going on, which I do like, but it's also important to keep in mind, like, what is the point of the deck? Is it a pitch deck that you send over for review, or is this something that you have on a projector while you're actually presenting over it? Because if it's the latter, then I would probably prefer Astros because it's a bit simpler and it's less distracting.
Whereas this one is a bit more like, I'm gonna send this over via email for them to read on their own time. So they're a bit different. I do like how Fable has the actual branding up top in every slide.
I think that that looks good. And Fable has some nice design elements with the way that it's structuring their columns and more variety in the slides, I guess you could say. So in my opinion, I think these are both good.
decks. This is also a pretty simple use case that you could probably get away with, you know, Sol, Terra, Sonnet. So Astra and Fable are probably overkill here.
So in my opinion, whichever one is cheaper is going to win this one. And it looks like Astra was much, much cheaper. Fable was 4 .8x the cost.
You can take a look at the time. You can take a look at the tokens here, but the API cost would have been about two bucks with Astra and about $9 with Fable 5 .2. Real quick, guys, I've got this completely free website design skill that I'm giving away.
I use this for every single website that I build, including our AI Automation Society site, which I think is pretty slick, and I absolutely love this website. It's called Scrollcraft, and it understands things like scroll -driven animations and layering, but also it has things in there like taste, typography, spacing, depth.
It's just a really good website design skill in general, and it's completely free, so the link for this will be down in the description, but let's get back to the video. All right, so number three, this is one of my favorite little benchmarks where I give these models an image of me and I have Canva open and I want them to paint me.
So same prompt, same thing. Let's take a look at these two outputs. So right here you can see this is Fable's output.
It's not too bad. You can see that there's some holes and you can see the marks and it's not super, super realistic, but it's definitely not bad at all. And let me pull up the actual real picture for you guys.
So here you go, just for reference. This is what Fable was working off of. You can see like the background, the book text, my hair, all that.
It's really not too bad, but I've seen better. Especially when we go over to Astros right here, which is just much, much better. As you can see in Canva, this is what it ended up looking like.
It is a much, much better output. It looks much more realistic. It definitely looks more like this real image than the other one does, as you can see.
I think it did really good with the shadows and the shading and that face looks like me as well, the text. And Astra always seems to crush cloud code out of the water with anything computer or browser use and a lot of the vision, things like that. Now what I thought was really interesting here is that Codex hit a stroke limit on Canva.
And what it did was it decided to export that, put it back in, and then on a fresh one, start painting over it again to get rid of that stroke limit, to sort of, you know, a loophole. But Fable kept running into that stroke limit. and didn't do anything about it, which is like, oh, I have to stop here then.
So it's interesting that Astra's reasoning was able to come up with a solution, whereas Fable wasn't. Especially when so far I found that Fable kind of over -engineers things and sometimes overthinks. So you would have thought that it would have made, you know, found a solution to that, but it didn't.
And look at this. This is absolutely disgusting. Fable ran for three hours, Astra ran for an hour and a half.
I think they both took longer than they should have, but either way, Fable ran way too long and it spent $63 via API billing. And Astra was about seven and a half. So Astra, no doubt, takes this one.
So right now, we are actually at three to zero. Astra is in the lead over Fable 5 .2. Okay, number four, what I did is I gave them a bunch of the actual videos from our AIS live event that was in July.
So the first one we did. And I wanted it to make an energetic 60 -second recap. And I've done this in the past before with different models, and they've been really, really good.
In fact, if you go to AIS .live, the hero video right there, that was built with Opus 5 .5 with one prompt. So very, very good results. So let me go ahead and play Asher's version real quick first, and then I will play Fable's version.
Okay, so I'm not happy with those at all. I think they were both really bad, and that's why I didn't play the full 60 seconds for you guys, because those, in my opinion, were both very, very bad. I've gotten way better outputs in the past.
I'm not sure if it's because of the way I prompted or what, but it did not do what I wanted to do at all. Like, not energetic, not cool. It didn't have any real speech.
Just very bad outputs, in my opinion, on both of these. Now, the lesser of the two evils, I would say, is Fable's version. So I'm going to give this one to Fable, I think, regardless of the cost here.
Even though Astra was about $4 and Fable was $15, I'm going to give this one to Fable just because it was better. But I really think that these were both pretty bad outputs. I think Opus 5 .5 crushes both of these.
Or maybe it was a prompting error, but I don't know. Those are both pretty bad. And we've seen AI video editing do things much better than that.
All right, so this one should be pretty interesting. This one, I did something similar. I gave them all of the recordings and assets from our other AIS Live events, so the one that we just did, and I asked for a 3D immersive playable world as if we were at a real virtual summit.
So let's open up the two worlds and see what we got. Okay, so here's Fable 5 .2s. I have to enter my name so I can get my badge.
So I'm gonna enter the event, and now you can see that... I guess this is a first person version, so I'm able to walk around. I can see all the speakers on the wall, which is pretty cool.
This is kind of inverted, which I don't love.
The mouse control is inverted, but yeah, I don't love how the mouse control is inverted, but besides that, not too bad. We can enter the scale track, where we can actually take a seat. and we're able to watch.
It says that this one starts in two hours and 20 minutes because it's actually holding a time right here, you can see, and I think it's gonna be live. So if I go to the agenda, we can actually see all of the real time. So this is pretty cool.
It was able to find that and create the agenda. This all feels on brand as well. So now it says here's something starting at the main stage.
I can click go there and it just takes me there. That's awesome. That was me.
Sipping my coffee. I'm not sure if you guys have read my book. Okay, not bad.
I'm going to turn the sound off just because it was probably overwhelming. But that's the real recording, and I could sit down and watch the session. So that's kind of cool.
I'm also able to, right from here, when I'm in the stage, I can pause it or I can zoom forward or full screen or pull it up in frame. I'm able to come out here and look at the agenda right there. I'm able to look at the sponsors.
I'm able to look at speakers. Not too bad. I think, you know, some of this stuff could have been designed a bit better because this looks very, I guess, cheap, but it's really not too bad and it's functioning as I would want it to.
We've got our booths in here, Glido, AIS +, Fish Audio, Kylon. We've got the booths right there. I like how we can just actually navigate though.
We can skip between different sessions. We can move around on the agenda. We can go to the sessions.
I think this is a really well done little creation. This is the VIP lounge. We got Russ in there.
Very cool. And here is... Astra's version which looks a little bit more like a concert.
It's kind of outdoor as far as Looks this one looks better like these avatars look better their shadows, but let's see how it actually functions So I'm able to take my seat I'm gonna mute that again, but you can see I'm taking a sip of my coffee. I'm able to scroll through this session, which was about an hour long.
I can't believe I just sat there and yapped for an hour. That was crazy. But then I can walk the floor, so I can walk around.
I can also return to my seat. I can pull up the program, so I'm able to see. Okay, I mean, this isn't bad, right?
I think I like the experience of the other one better. I think this one looks a little bit better, but I think the experience of the other one felt more on brand for AIS Live. Like even here, that logo is a bit wrong.
And I think the experience of the other one was better, even though this one looks a little bit better. So right now I want to give this one to Fable 5 .2. Let's take a look at the cost here.
We have similar time, but Fable once again is more expensive. And I'm assuming that's going to be a theme for the rest of this video. And what's interesting is look at the pricing for these models.
Fable 5 .2 is $8 per million input tokens and $40 per million output tokens, which is cheaper than Fable 5 .1. And GBT 6 .2 Astra is also cheaper than GBT 6 Astra, and it comes in at $8 for input and $40 for output. So they're actually priced the exact same, which is why you're seeing that Codex and Astra 6 .2 is just much more token efficient than...
the other model, Fable 5 .2, even though they actually are billed the same for those tokens, it's just a way more efficient model and harness at the moment. So I just wanted to point that out as well. But anyways, like I said, I'm going to give this one to Fable 5 .2.
So I believe right now we're sitting at three to two Ashes in the lead. All right, for number six, we have the Instagram Reels. These both turned out really, really good.
So let me play them for you and then we will give them a score. Well, it's a very sad day in the AI community because Anthropic and OpenAI have both admitted their models got worse after launch. Anthropic published post -mortems saying that users felt that Cloud Code was less intelligent and their own tests actually missed it.
And then Sam Altman said after GPT -5 dropped that it seemed way dumber the day after it launched. Well, it's a very sad day in the AI community because Anthropic and OpenAI have both admitted their models got worse after launch. Anthropic published post -mortems saying that users felt that Cloud Code was less intelligent, and their own tests actually missed it.
And then Sam Altman said after GPT -5 dropped that it seemed way dumber the day after it launched. So after you guys saw those outputs, which one did you think was better? I thought that Fable 5 .2 was much better when it comes to the sound design and even the selection of B -roll.
Fable 5 .2 was more engaging in my opinion, and it did cost, once again, about four times more. It was more expensive because it was less token efficient and used more tokens. But in this case, I think that it was worth it because I thought the output was much better.
So I'm gonna give this one to Fable 5 .2. All right, the number seven, I asked it to edit a YouTube video for me. So a long form YouTube video that was a talking head originally, and I wanted it to edit that using our style.
So I shot off this prompt. Let's take a look at. just a few seconds or a bit of each of the outputs.
Okay, so here's Asha's version. I have muted it just because it's gonna get too distracting. You can see that it's pulling in B -roll and stuff, but this is not the style that you've ever probably seen on a YouTube video of mine.
It literally just invented this style, which I'm not sure why it did that. I'm not happy about that at all. It looks pretty cheap.
I don't like this one bit, and I literally have a whole skill and context built around this. So this is just one of those situations, again, where for some reason Asha just completely ignored. what I wanted to do.
And I literally came in here and said, hey, why didn't you use my skill? Is it because it's not in the right spot or whatever? And it said, no, your skill's right here.
Codex just didn't use it. So I just have no idea what it did here. It used an old skill, which was a different one called YouTube Raw Full Pipeline.
So maybe it's my fault for not removing that skill. But all of the recent context and all of the previous folders, which it could have looked at, used this one. So I'm not sure what.
what happened there. So as long as Fable used the right skill, it's probably going to win this example. And here is Fable 5 .2's version.
I'm going to play the first intro section with sound. So this post got almost 48 ,000 likes and it's just Claude saying, you've used 90 % of your session limit and we've all felt that one before. because one developer burned 76 % of his weekly limit in less than a day.
So in this video, I'm gonna go over five fixes that are backed by Anthropix docs and people who actually have measured and tested their usage and my own logs, of course. I'm also gonna show you the viral token saver that made cloud way more expensive when it actually got... Yeah, so I mean, this is definitely better.
Look at the open loop it's creating. Look at these beautiful animations. The intro was Anthropix branded.
I love that. This used our actual skill. This used the correct thing.
This is exactly how I want the YouTube long form to be. These sorts of colors, these... sorts of animations, all of that sort of stuff in here.
Just much, much better. So obviously Fable 5 .2 is going to win this round. But look at how crazy this is.
Fable 5 .1, or sorry, Fable 5 .2 wins this round regardless. But look at this cost. This ran for over five hours, which is nuts.
Normally my YouTube videos do not take that long, and I normally do it with OPS 5 .5. Yes, it's a really beautiful output, but five hours is nuts. When Astra gave us that terrible output for seven bucks.
So this was expensive. Fable 5 .2 is an expensive model because it uses so many tokens, but this was a really good output that I. really enjoyed.
Alright, so now for number eight, I asked them to turn my GrokBot tutorial into a visual learning guide with screenshots and explanations and timestamps, and I gave it my actual full YouTube video. So let's take a look at these two learning guides. Alright, so this one is Fable 5 .2's output.
We're able to play the tutorial, which actually down here you can see plays the full video in this HTML, which is pretty cool, but I'm gonna close that. We have a map, so part one is We can actually click on these timestamps and it will take us to that section in the video, which is pretty cool.
We have what the bot is, which we can play these clips as well. So once again, it pulls up the exact clip. You can see down there it's muted.
And we have screenshots that have annotated captions, or not captions, annotations. So number one, you can see that's the bot's name, job label. Number two is the pin bots.
Number three is each bot has its own conversation. So it's obviously taking these screenshots, cropping them, and then it's putting these little annotations on them. bought descriptions, bought templates, and you can see that's basically how this whole thing goes.
It does a really nice job of showing us the different little tips as well as making the crops that we need in order to actually get a feel like we're watching this full tutorial rather than just reading some sort of guide. Once again, I really like how it has actually structured this, and I like how we're able to mark things off as got it, so then we're able to see, okay, cool, we've done, you know, two out of 26, and we could work our way through this learning guide.
I think this would be a nice, you know, hands -on learning experience. Now let's take a look at Astra's. We can see here, it's a bit different.
So if I click play tutorial, it pulls up the tutorial, as you can see, but it doesn't do it in the corner, so we're kind of stuck in this view. So we'll see if that's a good thing or a bad thing. Once again, it has little timestamps.
So like for the bot profile here, I can click on play and it just pulls it up full screen. If I go to description, we can see these different things. We're once again doing the annotations.
We have the workspace. We have, okay, interesting. So this is more like part one is up here and we go through one through five.
And then part two is down here and we go through six through 10. These annotations aren't as detailed and they're a bit smaller. But if I click on play it just pulls up the exact frame there So I you know, I think this is a little bit.
I mean all of these are pretty subjective, right? But anyways, I do think that this field guide is a bit cleaner and feels a little bit less Overwhelming even though I wish this was kind of fixed in the corner instead and then you can make it larger But I am gonna give this one to Astra And especially if we look at the cost here, you can see that Astra was about three bucks and Fable was about 17 bucks.
So this one's going to go to Astra 6 .2. Okay, now we have for number nine, trip planning. So I asked it to plan a trip.
I didn't really give it much direction. And you'll notice in most of these prompts, I was very vague. And I think being vague is really interesting to see what the models and the harnesses do.
So anyways, let's open up this first one, which was Astra. we can see we have the western passage we've got a little bit of three map here san francisco las vegas zion uh bryce canyon yosemite and we can see that we have different nights that we would stay in them now let's see what is actually planned for us is it just showing us the times and it's not really giving us any It's not really giving us things to exactly book.
Usually when I wanted to do this, I would love if it would give me like the hotels and flight options and all that sort of stuff already, which it didn't do in this one. So it kind of just gave us a really rough overview, but not really much. Like I would have to reprompt it to give us certain hotels, certain flights, and certain things to do.
But it looks pretty nice. And here is... the other one from fable that we're loading up it doesn't look as nice from like a design perspective it also has this weird flat map and only that section of the united states is highlighted so i don't know about that it's a little bit more overwhelming but we have san francisco yosemite zion so they both decided on doing the same sort of trip which is very very interesting because they they didn't work together on this anyways we have the different days and we can see fly west duration show this leg on the map And we can see as we go to these different stops how it sort of zooms in and shows us more of a schedule.
So it looks like it did a little bit more research because it's actually giving us some links that we could check out. Then it's having us drive from here to here. And the map is giving us a bit of a better experience there.
So, I mean, this is definitely a better working version of this actual trip planning compared to Astra's version, even though Astra's version looks much better. So it kind of seems like Astra has a really good feel for... I don't know, like the experience as far as aesthetically and Fable has a better feel for the experience as far as like the logic and the actual journey, which is an interesting theme to be able to identify.
But anyways, I would give this one to Fable because what I care about more when I'm trip planning is like the details it's giving me rather than just what it looks like. Even though this is more expensive, 18 .6 bucks compared to 2 .7 bucks. But this one's gonna go to Fable.
And I'm not gonna give you guys the running score anymore. We'll look at the running score at the end of how many I gave to Fable, how many I gave to Astra. All right, next one, we asked to turn this article into an engaging eight -slide AIS Instagram carousel.
And we have the brand guidelines around this already. So if I open up this sheet, oh man, not good at all. It did not use our guidelines at all.
This is not what they're supposed to look like. This should have been able to easily find our skill and find our context. make a carousel this is not what they're supposed to look like let's see if astra could have done it correctly and nope so both of these did not do it how they should have at all which i'm not very happy with really at all This is what they're supposed to look like.
We've been running this Instagram account with Claude, and it's pretty much been Opus that's done all of this and automated all of this, and they're all supposed to look like this. For example, you can see that they have this branding, and then as we swipe, we have these sorts of slides with the logo at the bottom, and then they always end with, you know, AI moves fast, so we share what's important, follow us.
And none of Fable or Astra did that, which is really, really interesting because normally... When I want them to do this, first of all, they're on a routine. But second of all, I literally will just send them like an X article or whatever it is and say, make a carousel.
I won't even say AIS Instagram and it does it. So not exactly sure why both of these models couldn't find our skill to do this. But that just reminds you that models are different and you have to make sure that your skills work with all your models because they always interpret everything a little bit differently.
So I don't like that. But once again, if we think the lesser of two evils here, which one do we go with? I would say we go with Astra.
I think that Astra's is a bit less overwhelming and just a better design, which isn't surprising based on the themes we've seen so far. And Astra also came in cheaper, which is great. So this round is going to go to our good pal, Astra 6 .2.
All right, so now what I did is I asked for an Excel workbook and an interactive dashboard for this data. So if we real quick just look at the interactive dashboard, they're very, very simple. It just shows a dashboard of the actual.
Revenue, and if I open up the dashboard in Cloud Code, you can see it's basically the exact same thing except for Astra's a bit more visual. So like I said, Astra's is more visual. It's a bit more aesthetically pleasing, but let's look at the actual workbooks.
So this one is Astra's, and you can see here that right away we have a financial overview. We've got some big numbers up top. We've got some very minimal charts.
We can see history. We can see forecast. It's very, very simple.
Nothing is structured too nicely. I'd be interested to see if these have formulas, which yes, they do. So far, they have formulas, which is great.
Because if they didn't have formulas, that would suck. So all of these do seem to have formulas, which is awesome. And if we go to Fable's version, it's a bit more colorful.
I'd say that the graphs and the numbers are probably more helpful. There's definitely more of them. We can see we've got drivers that's more detailed, more structured with different highlighting and colors.
We do have formulas in here, which is good. Let's make sure that we have formulas in most of the other parts of the doc, which we seem to, that also reference other sheets and things like that, other tabs. This document seems a bit more detailed, you know, and I think whether that's something that you like or something that you don't like, ultimately this one did give us more logic and reasoning to the actual visualizations and the numbers.
So I would say that Fable wins this sort of financial forecasting and modeling use case. If we look at the actual cost, you can see here that Fable ran for about twice as long, almost three times as long, and also cost a lot more. So 15 bucks compared to two and a half bucks.
I'm still going to give this one to Fable 5 .2, just purely based on the output. But you know, the thing is, most of these that I am giving to Fable, even though they cost a lot more, what I would probably want to do is I would want to build a skill around it and then just have Ash to do it because it's going to be able to do it for much, much cheaper and probably start to get the same outputs, especially once you are giving more specific prompts and skills.
So keep something like that in mind as well. Now, finally, we have a beautiful browser game where I am a pair of underwear trying to escape a bedroom. And I'm not going to exactly tell you which one is which yet.
This is the first one. So let's go ahead and play it. Brief escape.
I'm able to walk around like this and kind of change the angle a little bit. But for the most part, we're just trying to, I guess, jump around and escape here. It's a little bit slow.
Can I run? I don't think I can sprint. Pins ahead, jump or move around them.
So those pins would probably kill me. There we go, I reset. So I just have to get to the end.
This is a little bit boring, it's a little bit slow, and it's two -dimensional. Now let's take a look at this version. We have Brief Escape once again.
You know, it's so interesting. They're coming up with similar names. That just shows you like these AI models, like what their creativity is really like.
And it's hard. They're not very creative. They're taught to not think outside the box, right?
They're taught to recognize patterns. This one's definitely a bit better with the physics. It is, you know, 3D.
There's a cat here. And it's a bit more of a, like what I was looking for when I thought about like, you know, hey, can I build this little... browser game i did this previously in my last test with pistachios okay i'm oh wow i can like fly that's kind of cool you got a checkpoint here with one sock anyways that's what the game looks like so this one right here was fable 5 .2 and the other one was astra and fable was definitely a better gaming experience like i think a kid could actually play that and have fun whereas the first one was just kind of boring even though fable costs more here 19 compared to three fable is going to win this one So anyways, that brings us to the end where we've now looked over these 12 examples and we are one, two, three, four, five, six, seven.
Seven to Fable and five to Astra. But let's take a look at the actual final stats because this is gonna be really interesting. So real quick, cost by task.
Orange is Fable and blue is Astra. Just look at that. Look how much cheaper this all was.
This is nuts to me. And both of these models were running on high. Both of these models cost the same via API billing.
It's just that Codex with Astro 6 .2 is significantly more token efficient. Now, token efficient doesn't always mean better, but it is much more token efficient and it will spend less of your money than Cloud Code would with Fable 5 .2. And now if we look at some of the other totals, like obviously total API cost in total was 7 .2x more with Fable.
The runtime was also a lot more, almost 5x more, as you can see. Input tokens and output tokens obviously are going to be more with Fable. So Astra is faster, it's more efficient, it's cheaper, but the outputs aren't always better.
But it does hold its own, especially like I said earlier, if you really start to stack some skills, given that it actually calls the skill in the right time, that was absurd. But given all of that, if you do train Astra up in the right way, it will be... Really, really good and probably really, really efficient.
Now, none of these use cases were hardcore coding because I don't do a ton of hardcore software refactoring and big code bases and whatever. And sorry if that's something that you wanted to see. But when I do these tests, I like to test them on things that I actually would throw at these models so that I can give you guys my actual opinions on how they feel to me.
Because every model is going to feel different to different people based on their use cases. So anyways, hopefully this one was helpful for you guys. And if you guys did enjoy this one and you learned something new, then please give it a like.
It helps me out a ton. And as always, I appreciate you guys making it to the end of the video. And I'll see you on the next one.
Thanks, everyone.
The Hook

The bait, then the rug-pull.

Twelve real tasks, identical prompts, separate work trees, both models on high effort. The surprise is not who wins most. It is that the winner costs about seven times more at the exact same token price.

Frameworks

Named ideas worth stealing.

00:18model

Same-prompt head-to-head test rig

  1. Identical /goal prompt
  2. Separate work trees
  3. Both on high effort
  4. Track cost, tokens, runtime
  5. Blind-review outputs before revealing the model

A repeatable way to compare coding agents on real work by controlling prompt, effort, and isolation and then logging cost and time for every run.

Steal forany model bake-off
CTA Breakdown

How they asked for the click.

MENTIONED ON CAMERA
Storyboard

Visual structure at a glance.

intro
hookintro00:00
slide deck test
valueslide deck test03:11
Canva painting
valueCanva painting06:01
3D summit
value3D summit12:46
learning guide
valuelearning guide18:20
final totals
payofffinal totals28:42
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

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

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

More from this channel + related breakdowns.