08:41Nate Herk | AI Automation · Review
Claude Mythos is Finally Here
A screen-share walkthrough of Anthropic's dual model drop: Fable 5 for everyone, Mythos 5 for Glasswing partners only -- and why the host saw it coming.
June 9thThe argument in one line.
GPT-5.6 Sol and Claude Fable 5 score nearly identically in raw capability, but Sol wins on cost and reliability by a wide margin, while Fable's higher-quality output comes with roughly a 1-in-5 chance of silently failing to answer at all.
Read if. Skip if.
READ IF YOU ARE…
- You're building AI agents or automations and need to decide which model to route requests to for cost efficiency.
- You use Claude Code or Codex day-to-day and want a hands-on comparison of the two harnesses instead of a benchmark chart.
- You're deciding between a cheaper, faster model and a more expensive, more capable one for production API calls.
SKIP IF…
- You want a benchmark-only comparison — this is entirely hands-on testing, not a breakdown of published benchmark charts.
- You need model comparisons for tasks other than coding, agentic builds, and creative one-shot writing — that's all this video tests.
The full version, fast.
The video pits OpenAI's GPT-5.6 Sol against Anthropic's Claude Fable 5 across three side-by-side agentic builds (a browser game, an interactive website, and five visual mini-experiences) plus 130 quick one-off API calls scored for accuracy. Fable produced the more polished, creative output in two of three agentic builds, but took 2-3x longer and cost up to 20x more per run. Over quick API calls, Sol won 24 contests to Fable's 3, mostly because Fable silently failed to answer roughly 1 in 5 times; when Fable did answer, quality was a near-tie (0.966 vs 0.982). The aggregate scoreboard: Sol costs $16 total versus Fable's $63 for the same workload. The conclusion: treat Sol as the default workhorse for cost-sensitive and API-driven work, and reserve Fable for judgment calls, creative work, and the hardest engineering builds where quality matters more than price.
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 →Where the time goes.
00:00 – 00:51
01 · Intro & Benchmarks
Nate reacts to the GPT-5.6 launch page and its benchmark claims, then sets up the plan: skip the benchmark talk and run both models on real work all day.
00:51 – 03:31
02 · Test 1: Browser Bike Game
Identical prompt to Claude Code (Fable) and Codex (Sol) to build a playable open-world bike game. Fable's build wins on feel; Fable costs $14.22 with ~90k output tokens vs Sol's $4.50 with ~31k tokens.
03:31 – 06:16
03 · Test 2: Scroll-Stopping Website
Both models get creative freedom to build an impressive interactive scroll website. Fable's 'Ten Billion Years' cosmic scroll wins; costs $19.24 (80k tokens, 23 min) vs Sol's ~$1 (20k tokens, ~7 min).
06:16 – 12:38
04 · Test 3: Five Visual Worlds
Open-ended prompt for five unrelated visual experiences each. Sol wins this round — faster (7 min vs 15 min) and far cheaper (~$1 vs ~$15) — while Fable's set has visible bugs (stuck menu, dead inputs).
12:38 – 17:00
05 · API One-Off Tests
130 quick, stateless one-shot API calls scored for accuracy. Sol wins 24-3 largely because Fable silently fails to answer ~22% of calls; when Fable does answer, scores are nearly tied (0.966 vs 0.982). Aggregate: Sol $16 total spend vs Fable $63.
17:00 – 19:20
06 · Final Verdict: Manager vs Worker
Nate lays out his mental model: Fable as the creative, judgment-heavy 'manager'; Sol as the fast, cheap, reliable 'worker' — each has a clear lane rather than one outright winner.
19:20 – 19:50
07 · Where Each Model Ranks
Nate places four current coding models on a single capability line: GPT-5.5 and Opus 4.8 roughly tied, Sol clearly ahead, Fable further ahead still — with a real gap remaining between Sol and Fable.
19:50 – 20:25
08 · Final Thoughts
Nate closes by reminding viewers that model choice is use-case and workflow dependent, not a single universal answer, and asks for a like if the breakdown was useful.
Lines worth screenshotting.
- GPT-5.6 Sol and Claude Fable 5 score within 0.016 points of each other on raw capability (0.982 vs 0.966), but Sol costs a fraction as much per correct answer.
- Across 130 quick API tasks, Fable failed to answer roughly 1 in 5 times, while Sol delivered an answer 100% of the time.
- On an identical agentic build, Fable's output used nearly 3x the tokens of Sol's (about 90,000 vs 31,000) for the same browser-game prompt.
- Building the same interactive website cost $19.24 with Fable and about $1 with Sol — roughly a 20x price gap for comparable output.
- In one test, Fable took 21-23 minutes on agentic builds that Sol finished in as little as 7 minutes on the same prompt.
- In head-to-head one-off API calls, Sol cost $16 total while Fable cost $63 for the same 130 tasks — nearly a 4x gap.
- The creator ranks four current coding models on one line: GPT-5.5 and Opus 4.8 roughly tied, then Sol clearly ahead of both, then Fable further ahead still.
- Despite Fable's higher price and slower speed, the creator still prefers it for knowledge work, creative brainstorming, and avoiding over-engineered solutions.
- Sol's cost advantage isn't just a lower price per token — it also uses meaningfully fewer tokens to reach a comparable answer, compounding the savings.
- The creator frames the choice by role rather than raw score: Fable is the 'manager' worth paying for on judgment calls, Sol is the 'worker' worth defaulting to for volume.
Fable is the manager, Sol is the worker — pick based on the job.
MODEL SELECTION
When two AI models are nearly tied on raw capability, the deciding factor is reliability and cost per correct answer, not the benchmark chart.
01Intro & Benchmarks
- Benchmark charts can show one model 'blowing another out of the water,' but real-world task performance often tells a very different story than the published numbers.
- When two models look close on paper, the only way to tell them apart is running your actual day-to-day work through both, not reading the marketing page.
02Test 1: Browser Bike Game
- Identical prompts to two different models can produce two working results with a genuinely different feel — evaluate them blind, without knowing which model made which, to avoid bias toward a favorite brand.
- The pricier model can output nearly 3x more tokens for the same task, and that extra output doesn't always buy a better result.
03Test 2: Scroll-Stopping Website
- A vague, creative-freedom prompt is a good stress test for a model's taste and originality, not just its ability to follow instructions.
- The cost gap between models can widen dramatically on more ambitious asks — the same class of task can cost roughly 20x more with one model than another.
04Test 3: Five Visual Worlds
- Even the 'losing' model in a head-to-head can produce work with real bugs; read past the polish to check whether the thing actually functions.
- Don't assume more expensive always means better — the fastest, cheapest option won at least one of the three head-to-head builds outright.
05API One-Off Tests
- For quick, stateless one-off questions, a lower-capability-but-reliable model can beat a rarely-wrong-but-often-silent one; reliability of delivery matters as much as peak quality.
- A model that fails to answer roughly 1 in 5 times effectively lowers its own average score, even if its answers are excellent when it does respond.
- When cost and quality are close, ask what capability level the task actually requires — a task that needs a five-out-of-ten model doesn't need a ten-out-of-ten one.
06Final Verdict: Manager vs Worker
- A useful mental model for picking between two capable AI tools: one plays 'manager' (judgment, creativity, pushback), the other plays 'worker' (speed, price, execution) — route tasks by role, not just by benchmark score.
- The more expensive, more careful model in this test was also the more conservative one — pointing out edge cases and pushing back rather than immediately executing — which is a tradeoff worth naming explicitly, not just tolerating.
07Where Each Model Ranks
- Ranking models on a single line, from weakest to strongest, is a simple way to communicate a felt sense of a fast-moving field — even if it's subjective and will look outdated in months.
08Final Thoughts
- The right model choice depends entirely on the use case, your own workflow, and how your tools are set up — there's no universal 'best' model, only a best fit for the job in front of you.
Terms worth knowing.
- Agentic build
- A task where the model runs autonomously inside a coding harness (like Claude Code or Codex) — writing, testing, and iterating on code over many steps — rather than answering a single one-off prompt.
- Token efficiency
- How many tokens (units of text) a model uses to produce its output; a more token-efficient model reaches a comparable result while billing for less text, lowering cost even before accounting for price per token.
- Silent refusal
- When a model fails to produce any usable answer to a prompt — through a timeout, an empty response, or an internal guardrail — without giving an explicit error explaining why.
- Harness
- The surrounding tool, such as Claude Code or Codex, that wraps a raw AI model with the ability to read files, run commands, and make multi-step edits, rather than just answer chat messages.
Things they pointed at.
00:00toolGPT-5.6 (OpenAI)
00:00toolClaude Fable 5 (Anthropic)
00:51toolClaude Code
00:51toolCodex
Lines you could clip.
17:00
“Overall, I feel like Fable is a better manager, and that's how I think of it. I think of it as a cofounder. Think of it as a manager. Soul is just a really, really good worker.”
12:55
“The record here... is that Soul won 24 times and Fable won three times... Soul costed $16 while Fable costed $63.”
18:52
“Fable five is in another tier. It's just on another level, and Soul is really good.”
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.
metaphor
00:00So we finally have GBT 5.6 Soul, Terra, and Luna. And the benchmarks are insanely impressive because it basically shows that Soul blows Opus and Fable out of the water and is significantly cheaper. So I'm gonna quickly flash a few of these benchmarks on screen right now, but but I don't wanna just talk about the benchmarks.
00:17I think that's extremely boring. I've been playing around with these models all day long. I've been putting them side by side, meaning g p d 5.6 and Fable five, and I've been driving my day to day work with these two models so I can actually tell you which model do I like for which use cases and why.
00:30So don't wanna waste any of time. Let's just get straight into today's video. Now real quick before we hop in, I did wanna call out that I do think that this landing page is pretty awesome.
00:37I love the soul and the terra and the luna back here, and you can see that there's little fives and sixes kind of like drifting all over and inside of these planets as well, which I thought was a nice touch. But anyways, let's get into what I actually did today.
00:51Alright. So I'm just gonna walk you guys through some of the experiments I did, and then at the end, I'm gonna actually talk about, like, the pros and cons of each and my overall consensus. So the first thing I did, I have Claude on the left here with Fable five.
01:01I've got Codex on the right here with GBD 5.6 Soul, and I gave them both the exact same prompts. So I gave them both the slash goal, and I told them to build a genuinely fun, playable, open world byte game that runs in the browser.
01:13I gave them these other rules. I'm not gonna read off this whole prompts, if you want to take a look, feel free. I gave them creative freedom, and I just wanted to see what they could come up with.
01:21Okay. So I haven't played any of these yet, but I'm not gonna tell you which one's which. We're gonna see them both, and then I'll reveal.
01:26So w a s d to steer, space to bunny hop, q and e for air tricks, shift for boost. Okay. Let's go ahead and drop in.
01:33Nice. Okay. So we can move around.
01:35We have a city here. The camera's a little bit hard to maneuver with. You can't drive through the building, so that's good to know.
01:42Okay. Just got a little trick in the air there. So this is honestly working pretty well.
01:46I don't love the actual way that this works. Like, the top down bird's eye view, don't think I like.
01:52It's a little bit hard to control. But, hey, functionality wise, the coins are working. The controls work.
01:57I just jumped over a building, and I can bunny hop. Okay.
02:00Cool. And it doesn't look too bad either considering this didn't take it very long. So that was the first game.
02:06Alright. And now here is the other version. Let's go ahead and see.
02:09It's pretty similar controls, so let's hop in here. Okay. So this one is more like a three d, almost more of like a POV sort of style.
02:17I can jump. I can spin. Oh, I don't know what just happened there.
02:21But I'm in the city. I've got the coin still. Let's see if I can go try to get on this ramp and see what the the physics of that look like.
02:28Wait. Am I on wrong side of the ramp? Yep.
02:30That was the wrong side of the ramp.
02:33But this definitely has more of, like, that I don't know. Like, more of a GTA vibe, like an open world game that you can actually, like, I don't know, explore. It feels like the map is much bigger.
02:42I mean, this is definitely a clear winner, so I think that we've decided that. Alright.
02:48So in this example, the second one, which is the one I thought was the winner here, was fable five. So that's on the left hand side here. This took twenty one minutes and thirty seven seconds, whereas the other side, GBT Soul took twenty three minutes.
03:01So similar on the timing here, but let's look at cost. So this Fable run would have costed me about $14.22, whereas the GBT Soul run would have costed 4 and a half dollars.
03:11Now what I think is really interesting here is the output tokens. Fable outputted about 90,000, and Soul only outputted about 31,000.
03:19So we've consistently seen that Codex and GBT are way more token efficient. And sometimes that's better and sometimes that's not. In this case, I would say that it wasn't better because Fable's output, in my opinion, was, like, not just a little bit better.
03:30It was significantly better. Alright. So let's move on to the second test.
03:34So this one was kinda similar. I just said, hey slash goal. Build me the most impressive interactive scroll stopping website that you can imagine.
03:40So I gave them complete creative freedom. I just wanted to see what the models did with that. So let me pull up both versions once again.
03:47Alright. So once again, I'm not gonna tell you which one is which. I want you to just try to form your own opinion.
03:51This one also has sound with it, so I'll make sure the sound is turned on. So what you can see here is ten billion years. As And we scroll down, we're gonna get, like, kind of a journey, but everything in the background is pretty interactive.
04:02So it's very, like, three d. So watch as I scroll down here. We start off with the cloud, not too much, but you can still see that there's, you know, like, responsiveness to my mouse movements.
04:12Then we have the collapse. And once again, we've got some cool, like, galaxy looking very three d. And, basically, that's the theme of this whole thing.
04:19It's telling the story, I guess, of, like, you know, how we got to where we are. This is a pretty cool effect. First light and the swelling, so the fuel starts to run out, and then we had a big boom.
04:33So that was a a big noise here for, um, basically, the supernova. This is a nice effect as well.
04:39And then like I said, we just kinda keep going through with this sort of feel up until the very end where we are just at us. So we have the word you. Pretty cool.
04:47So that is what one of the models decided to do. Let's take a look at the other one. We have Vesper archive.
04:54And as we go down, we have a similar sort of thing with sound. And it's kind of interesting because they have, like, a similar feel. Right?
05:00Like, a three d sort of background that's interactive with my mouse, and it's telling some sort of story. It's telling some sort of journey, but they're both not bad. So far, I definitely think that the first one was a little bit had more of a wow factor for sure.
05:13It was a bit more immersive. This one tells a nice story, and it still has some impressive features. But I think the first one so far was definitely on a different level.
05:22So which one was Fable? Fable was the first one. So once again, think that Fable did a better job here.
05:27What you can see is Fable took twenty three minutes, whereas SOL took almost seven minutes. And if we go down to the actual cost, Fable was $19.24 with 80,000 output tokens, and SOL was only a little over a dollar and output 20,000 tokens.
05:42So it's interesting. Like, if we had SOL working enough to spend as much as Fable did, would SOL's output have been just as good, if not better? Maybe.
05:51SOL is significantly cheaper on top of the fact that it typically spends less tokens in general just because of the harness and because of the way the model is trained, I guess. But, yeah, I mean, this was almost 20 times cheaper for Fable. But the question is, is Fable's output, was that 20 times better?
06:06And I don't necessarily think that it was. But just because we're going off of, like, the actual quality, I'm going to say that I would have taken Fable's output in this one as well.
06:16Alright. So I did one more side by side experiment. This one's interesting.
06:19I gave it a different goal prompt, I said, design me five fundamentally different visual elements. They can be games, presentations, websites, simulations, whatever you want. I really wanted to see once again, what does the model do with the super ambiguous prompt?
06:30So let me open up the different outputs right here, and we'll go ahead and compare. This time, I will go ahead and tell you which one is which as we're looking at them.
06:39Alright. So this one is Fable, and this one is Soul. So they each output one site, and then we've got five different kind of, like, simulations or sister sites to go through.
06:50So let's just start with Fable and go through all of them, and then we'll go to Soul. So this is what we got so far, the gallery, five self contained worlds. Let's look at singularity.
06:58So I don't really know what's going on here. It says move to warp space time, hold to feed it, I guess, and then you hit space to detonate.
07:06So if I hit space, we just blew something up, I guess. So I'm not really sure what they're going for here, but that's number one.
07:15We have second one, which is called Terra, where we kind of, like, have this flight simulator almost where we can steer with our mouse. We use w or s to speed up or speed down.
07:23We can go up or down. We can change the time of day. So as far as, like, the physics on this, it's not too bad.
07:29Everything is kind of rendering in a bit strange. I'm not sure if they were doing that intentionally.
07:33Maybe that's, like, the vibe of this, but pretty cool. You know, not too bad. I don't love it, but I mean, for the fact that Fable just kind of wanted to do this and did it, it's alright.
07:46Okay. We've got Orbit, which is an arcade game. Two rings, one input.
07:51So I think it looks like we just click to switch rings. We're trying to collect the coins.
07:56Okay. We're trying to avoid those things is what it looks like, and that's how you get shattered.
08:00So, I mean, it plays fine. It feels very responsive. I'm not sure.
08:04This is a bit of a bug, though. The menu is, like, taking over me actually playing.
08:09So I'm playing right now, but the menu isn't going away. So that's a bit of a bug.
08:14Let's go ahead to number four. We've got ink flow. Click for new piece.
08:20I can change the palette. I can save them as a PNG. I think this is just a weird like a board button sort of thing.
08:28I don't know if you guys ever did board button, but that's what this reminds me of, like, just some very random thing. And the last one is the descent. So scroll from sunlit surface to challenge okay.
08:38So this is kind of like a story again. It's like an interactive story where we're actually going underwater.
08:43The background's changing. We're seeing different creatures go through deeper than the Titanic. This pressure, the sunlight, the temperature, all of that's changing.
08:53So kind of just another little journey experience. So that was Fable. I will say nothing overwhelmingly impressive there.
09:00Right? Let's go to soul, which already I think this looks better. This honestly kind of looks like a claw design sort of output, but, anyways, impossible objects, five tiny worlds, no frameworks.
09:11The first one is Aurora Orchestra. So I guess we're just kinda playing here with the northern lights.
09:20Yeah. Not sure exactly what's happening, but we can change the voice. We can change the sky.
09:24We can hold for intensity. We've got other ones in here, so we can change, like, the color of the northern lights, I suppose. Not sure what's going on here.
09:33That might be a bit of a bug. It seems like three and four don't actually work, but two and oh, now two and one don't work at all either. Okay.
09:41So another little bug there. Let's go back to the gallery. Let's go to number two, Atlas of Lost Echoes.
09:47We've got a map. My cursor is showing where north is. What happens if I click on this?
09:53Sleeping sea. Okay.
09:57I'm not really sure what we're doing here. In the bottom right, can see it says echoes recovered. I guess I have to recover all five of these echoes.
10:04So let's see if anything happens.
10:07Every place that misses us, Atlas complete. Keep exploring. Okay.
10:10So not really sure what that is. Let's go ahead and go to the next one, the glyph heist. So the grid is hiding contraband words.
10:17Type every signal before the trace completes. Okay. So this is just like a typing game, I guess.
10:25Am I getting points? Oh, yeah. The points are in the top left.
10:30Woah. That's not even a word. Okay.
10:33So this is just like a typing game, I guess. I don't really see much, like, feedback. I'm just kind of typing, and those that's not a word either.
10:41I mean, I guess that's the point is it's trying to trip me up. Oops. Yeah.
10:44That's kind of the point, I guess. Okay. That's alright.
10:47Um, I mean, kind of a cool vibe. Right? Like, the actual aesthetics.
10:50I don't think you're losing lives, though. Like, I don't know how this ever ends. Also, I already had a high score in here.
10:58So definitely kinda buggy. Um, and let's go ahead and go to I guess we got two left. So this one this is just a slide deck, it looks like.
11:07It looks like there are five slides. Not bad, though. I actually think that these are designed pretty well.
11:13This is interesting, though, because this definitely looks like a claw design to me. Like, this definitely looks like something that Fable would whip up.
11:20I don't know why. The fonts and the color scheme just feels like a clawed thing. Anyways, the last one we got here is a tide pool.
11:26So life is waiting. Touch the water to begin. Okay.
11:30I just dropped in a bunch of seeds. We've got a population of 320, diversity of two.
11:35I can give them sunlight. I can give them fresh rain. So this is very much like a simulation.
11:41This one already feels like it's, like, slowing down my computer, so I don't know how much is going on back here, but not too bad. I like the ones that are simulations.
11:50I think they're kinda fun. So yeah.
11:53Honestly, I don't even know what to say here. Like, I don't know which one to say I'm gonna hand the crown to. I think I'm just gonna go with Soul for this one because I think, in general, it had a bit of a better idea of, like, diversity and experience.
12:06The interesting thing here is when we take a look at the speed and the cost. So I thought that this run would take the longest out of all of them because I was saying, hey. Five very different things.
12:14I thought it would take maybe five times as long as the earlier ones, but this one was actually, like, the quickest on both fronts, which I thought was really interesting. This one took fifteen minutes for Fable, and for Soul, this took seven minutes, which I thought was really, really fast.
12:27And on the cost side, this costed Fable about $15 with 65,000 output tokens. And on SOL, it took, once again, about a dollar and 22,000 output tokens. So very, very interesting to me.
12:38Alright. So those Agencik experiments with Cloud Code and with Codex, that was one thing.
12:44And that's typically how I'm gonna be actually driving these two models, so it was important for me to play around with stuff. And, of course, all day, I used it for just my general work. But I also wanted to try just explicitly over the API, just talking to the models.
12:56And a lot of these were kind of like a one shot, hey. What's this? Give me an answer.
13:00What's this? Give me an answer. Rather than like an agentic loop with multistep reasoning, but I wanted to see what it looked like for, you know, speed and cost as well.
13:08So the record here, which is really interesting, is that Soul won 24 times and Fable won three times. There were also a lot of times when Fable refused to answer, like, literally just wouldn't submit an answer. It refused, which I thought was interesting because, you know, there's the security guardrails are baked in pretty hard.
13:23But, also, look at the spend. I mean, Soul won 24 to three for these kind of quick one off tasks, and Soul costed $16 while Fable costed $63. But as far as, like, the actual capability, when Fable actually answered, it was pretty even.
13:36The reason why Soul dominated apparently is because Fable failed to answer a lot of the calls. But as you can see over here, score one answering, Soul had point nine eight, and Fable had point nine six six. So they are very similar models, and I think it's just a perfect example of, like, the unit economics.
13:52Based on this task, let's say the task is a five. Do you need a model that is a 10? Probably not.
13:57You only need a model that's maybe, like, a five or a six. So in this case, with these quick one shot, you know, API requests, probably stateless, SOL seems to crush Fable here.
14:07And from an input output tokens perspective, soul is basically half the cost of fable. You know, half on the input and then a little over half on the output, but it's basically half. And if you compare soul to, like, opus 4.8, they're pretty much the same.
14:20Same on the input and then very similar on the output. So that's pretty interesting to me. And it honestly feels like a better head to head comparison would have been Soul 5.6 versus Opus 4.8 rather than comparing Soul 5.6 and Fable five.
14:33Like, I just think that they're not even on the same level as far as, like, a next gen model. Think about the fact that Fable jumps to a five. Soul, which is apparently their most capable model, why would it not have been Soul six?
14:43Like, why are we 5.6? Doesn't a model that should be so so good deserve a full step up if there's, you know, fundamentally something so different that makes it, like, a really powerful scary model. But they decided to only bump it up to Soul 5.6, which makes me wonder, like, what are they cooking back there, and what are they maybe waiting to drop as their next GPT six?
15:03Because I think that's when we have the next gen model that I truly think actually goes head to head with Fable because right now, I do not think that Soul does. I'll talk about that a little bit more once we get into the next section, but let's look at this. So score, obviously, Soul won here.
15:16But score when answered, obviously, it was much closer. So that's kinda like the reliability thing. Now on the cost side, the API calls, SOL was much cheaper, obviously, $9 to about $14.
15:27But then what's interesting is the gap here is not that big. Right? The gap here is about $5.
15:31And if you think percentage wise relative, it's even less, like a smaller percentage than how wide this gap is When we get to the actual agentic reasoning with Cloud Code as the harness and Codex as the harness, the gap in cost was just much, much bigger, which is really interesting. And overall, Fable is much more expensive than Soul.
15:47And I think the token thing is just so interesting because let's say that Fable and SOL costed the exact same from an input and output token perspective. SOL would have still been meaningfully cheaper because it just was more efficient with tokens. I don't know exactly why that is, but it just seems to be so much more efficient with tokens.
16:04It's probably has a lot to do with codecs too because when I've done codecs and cloud code comparisons in the past, codecs has always been faster and more efficient with tokens. But, anyways, just something that I think is pretty interesting. This is also interesting on the speed side.
16:16So over API, the median API latency from Soul was quicker than Fable, but the mean, so the average, was Fable being quicker.
16:26So that means, like, Soul is a little bit more all over the board. Some runs were super, super long. Some runs might have been really, really quick, Whereas Fable was more consistently just kind of hovering around, like, the 20 mark, and that's why its median was still higher than Soul, but its mean was lower than Soul, which I just think is pretty interesting.
16:43But then when we see the agentic builds, Fable was pretty much always going to be slower. So after running these different types of tests, some were fun, some were for design, some were, you know, just via API, and then, of course, my actual driving this on the day to day, here's where I landed.
17:00Overall, I feel like Fable is a better manager, and that's how I think of it.
17:06Soul is just a really, really good worker. Really good worker. Like, I think that having Fable orchestrate a bunch of Soul agents would be amazing.
17:13I gave them each their other things that I would say, like, they are better than the other end. So soul is obviously better on price. It's much better on computer use.
17:21It's much better at playing devil's advocate because it feels like it's just a little more careful at finding the bugs and looking at things from different angles. I also gave it the verification because of the computer use and because of this whole devil's advocate and just, like, really being careful and really thinking about what did they say, what did the user tell me to do, and what am I gonna do, and I'm gonna do exactly that.
17:39Whereas sometimes fable's a bit more like, it'll push back a little bit more, and it's almost like sometimes sassier. You know? Like, sometimes it's not just like I want to take a request and action it.
17:49I want to push back and be like a consultant here, which is good thing. Right? That's why I said it's a manager.
17:53There's pros and cons to that. And then, of course, I gave soul the speed. Now Fable, I said, it's just much more creative.
17:59It feels more creative. It feels like it's better at writing. It feels like it's better at brainstorming.
18:02It feels like it's better at advising. It's really strategic. The knowledge work all day, I still prefer Fable, but, you know, it's just not the the cost.
18:09I can't always justify it. It felt like it was better at making videos. If you saw, I made two videos recently on YouTube.
18:15One was called, like, fable five made this entire video, and the other one I just dropped earlier today, which was called soul made this entire video. So go watch both of those. They each created YouTube videos for me completely from scratch, and just see which one you thought was better.
18:27I liked fables better. I also said that fables better at not overengineering because Soul runs so many tests, which also is interesting when you think about the token thing.
18:36Right? Like, how can it be so efficient with tokens if it writes so many tests? And sometimes it will overthink.
18:41Especially when I had Soul on ultra, it was just going and going and going and going, which, you know, sometimes good, sometimes bad. There's always pros and cons, but just pure capability. I mean, can't even argue.
18:52. It's just on another level, and Soul is really good.
18:57I think it's more comparable to Opus 4.8. It's definitely better than GPT 5.5, but it just it's not a Fable five.
19:02And, honestly, I think if it was Fable five tier, it would be charged at Fable five tier. Like, it would be priced that high, and it's not.
19:10It's priced like Opus. So overall, what I'd say is Fable is good for my reasoning and my judging, and then Soul is just shipping and executing AI model. And just so we kind of, like, put it into perspective, I kind of tried to, like, map out where I think that these top four models that, you know, we're talking about right now sit.
19:28I've got GBT 5.5 and Opus just very, similar. I would say because I drive most of my work with Opus, I would put g b t 5.5 just, like, a little bit behind. Soul is definitely in front of both of these.
19:37Like, that's very clear. It's a very, very good model, but I still think there's a pretty large gap right here between, you know, Soul actually being to be on Fable's level.
19:47Like, there is still a pretty big gap here. So that's the way I feel about it. Hopefully, all of this makes sense to you guys.
19:52Like I said, it's definitely one thing to look at the benchmarks, but really what you have to do is you have to get your hands dirty. You have to do actual work that you do on your day to day, and it's a feel thing. Because some people will probably come out the gate and say soul is better than fable, hands down, and some people will not agree with that at all.
20:05It's all about your use case and your feel and the way that you instruct it and the way that you've got your harness and your skill set up. So everyone's a little different, but that is kind of my opinions after playing around with this model all day long. But as always, if you guys enjoyed the video or you learned something new, please give it a like.
20:19It helps me out a ton. And as always, I appreciate you guys making it to the end of the video. I'll see you on the next one.
20:23Thanks, everyone.
The Hook
The bait, then the rug-pull.
OpenAI just shipped GPT-5.6 Sol claiming to blow past Opus and Fable on the benchmarks — so Nate Herk spent an entire day running it against Anthropic's Claude Fable 5 on real work instead, before publishing his own cost-and-win-rate scoreboard.
Named ideas worth stealing.
17:00concept
Manager vs Worker model-selection framework
- Fable = 'manager': creativity, advising, knowledge work, judgment, not over-engineering, pure capability
- Sol = 'worker': price, computer use, devil's advocate/bug-finding, verification, speed
The creator's mental model for choosing between two AI models: route creative and judgment-heavy work to the more expensive, more capable model, and route high-volume execution to the cheaper, faster, more reliable one.
Steal forany workflow that has to decide which AI model to call for a given task based on a cost-vs-quality tradeoff
How they asked for the click.
VERBAL ASK
20:10subscribe
“if you guys enjoyed the video or you learned something new, please give it a like”
Soft end-card ask only — no in-video sponsor read or hard pitch; monetization runs through description links to Nate's community and affiliate tools instead.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
