The argument in one line.
GPT-5.6 Sol narrows the coding gap with Fable and pulls ahead on knowledge-work tasks, making a three-model workflow — Fable orchestrating, Sol handling knowledge work and mid-tier builds, Opus 4.8 filling gaps — the most cost-effective setup for someone using all three families.
Read if. Skip if.
READ IF YOU ARE…
- You regularly switch between multiple frontier AI models and want a current, task-by-task read on which one wins where.
- You're deciding whether to route agentic coding work, document generation, or creative writing to a specific model family.
- You want to understand the new 'effort level' and parallel sub-agent settings that OpenAI and Anthropic both shipped, and how they actually differ under the hood.
- You're comparing per-token pricing across a frontier model lineup before committing budget to one provider.
SKIP IF…
- You only use one model and have no interest in swapping providers.
- You want rigorous, reproducible benchmark methodology — this is one person's blind subjective scoring across ten prompts, not a controlled study.
The full version, fast.
OpenAI's GPT-5.6 launch ships three tiers — Sol (flagship), Terra (mid), and Luna (small) — plus a new 'ultra' effort level that coordinates four agents in parallel by default. The reviewer ran Sol against Fable, Opus 4.8, and GPT-5.5 across ten blind tasks: five browser builds (Rubik's cube solver, a 3D apartment reconstruction, a galaxy map, a shooter clone, a cinema knowledge graph) and five knowledge-work tasks (a graphic novel, dungeon-crawl writing, a voice-matched YouTube intro, a launch deck, a brand rebrand). Fable still wins on deep agentic coding, matching its higher SWE-bench Pro score, while Sol edges ahead on Terminal Bench and a long-horizon task benchmark. On knowledge work, Sol was the most consistent top performer of the four, and it costs less per token than Fable while landing close to Opus 4.8. The reviewer's practical takeaway: use Fable as the top-level orchestrator for hard coding problems, route knowledge work and lighter builds to Sol, and fill in with Opus 4.8 elsewhere.
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 – 00:54
01 · Cold open
GPT-5.6 launches as three tiers — Sol, Terra, Luna. Pat frames the video: ten blind tests against Fable, Opus 4.8, and GPT-5.5.
00:54 – 01:20
02 · Spinning up all the builds at once
Pat kicks off five parallel coding builds using a Claude-orchestrated skill that spawns five terminal windows at once, then lets them run in the background while he covers model differences.
01:20 – 02:41
03 · Sol, Terra & Luna explained
Sol is the flagship for ambitious agentic work; Terra performs like GPT-5.5 at roughly half the price; Luna is the small, high-volume tier that reportedly beats Opus 4.8 on agentic coding benchmarks despite being built for lightweight jobs.
02:41 – 05:24
04 · Reasoning effort & parallel agents
OpenAI adds two new effort levels beyond extra-high: max and ultra, the latter coordinating four agents in parallel by default. Pat contrasts this with Anthropic's dynamic workflows, which write an actual deterministic program that can scale to hundreds of sub-agents without hitting context limits.
05:24 – 07:38
05 · Cost & benchmarks
Pricing across the three GPT-5.6 tiers versus Anthropic's Haiku/Sonnet/Fable lineup, followed by four cited benchmarks: Terminal Bench 2.1 (Sol wins), a 50+ profession long-horizon benchmark (Sol wins by 12 points), SWE-bench Pro (Fable wins), and GDPval (Fable edges Sol).
07:38 – 08:32
06 · The ten tests (and viewer suggestions)
Pat lays out the full slate: five AI-coding builds and five knowledge-work tasks, several sourced from comments on a previous review video.
08:32 – 10:41
07 · Rubik's Cube solver
Four models build a browser-based, functionally solving Rubik's cube. GPT-5.6 Sol's first run failed to render; a re-run later in the video placed it second behind Fable, ahead of Opus 4.8 and GPT-5.5.
10:41 – 14:10
08 · Monica's apartment in 3D
Each model reconstructs the Friends apartment from a floor plan in Three.js. Opus 4.8 nailed the layout accuracy best; Sol's render looked polished but missed the actual room arrangement.
14:10 – 17:10
09 · Interactive Milky Way learning map
An explorable, clickable galaxy map pulling real astronomical data. Pat ranks Sol first here, ahead of GPT-5.5, Fable, and Opus 4.8 last.
17:10 – 19:08
10 · Counter-Strike 2 clone
A playable first-person shooter, benchmarked against a build Pat originally made with Fable 5. Fable wins again; Sol places second, ahead of Opus 4.8 and GPT-5.5, which had broken controls and no visible weapon.
19:08 – 23:05
11 · Cinema knowledge graph
Models parse public movie/actor datasets into an explorable knowledge graph with a 'six degrees' path tracer. GPT-5.5 wins on creative execution; Sol places second; Fable pulls in surprisingly little data and ranks last.
23:05 – 26:05
12 · Interactive graphic novel
A branching story with three endings and AI-generated art via GPT ImageGen. Sol produces the best-organized, best-illustrated output, ahead of Fable, Opus 4.8, and GPT-5.5.
26:05 – 28:14
13 · Text-adventure dungeon (writing test)
A 12-room dungeon crawl scored purely on prose quality. Sol is Pat's favorite output, with Opus 4.8 a close second; Fable ranks last on this writing-specific task.
28:14 – 31:18
14 · YouTube intro in the creator's voice
Each model writes a video title and hook from a corpus of Pat's past intros. GPT-5.5 matches his style closest, ahead of Sol, Opus 4.8, and Fable.
31:18 – 34:19
15 · PowerPoint on the GPT-5.6 launch
A real, editable launch deck. Sol is the only model to build a genuine PowerPoint structure and independently pick up the launch's planetary visual theme, ranking first ahead of Opus 4.8, Fable, and GPT-5.5.
34:19 – 38:13
16 · Spirit Airlines rebrand
A full 360-degree brand refresh — logo, copy, web design, ad mockups — for the bankrupt airline. GPT-5.5 produces the most realistic AI-generated imagery; Sol's web design is Pat's personal favorite; Fable ranks lowest on this task.
38:13 – 39:31
17 · The verdict: when to use which model
Fable still wins on AI coding overall; Sol wins on knowledge work and costs less than Fable while nearing Opus 4.8 pricing. Pat proposes a three-model workflow: Fable orchestrating, Sol on knowledge work and mid-tier builds, Opus 4.8 filling gaps.
39:31 – 40:03
18 · Outro
Pat asks for more test ideas in the comments and teases a follow-up video testing Sol and Fable on bigger, more creative builds.
Lines worth screenshotting.
- GPT-5.6 launched as three tiers — Sol (flagship), Terra (mid-tier), and Luna (small) — mirroring the low/mid/small structure other frontier labs use.
- Sol is priced at $5 per million input tokens and $30 per million output tokens, notably cheaper than Fable's $10 input / $50 output pricing.
- Terra performs roughly like the prior GPT-5.5 model but at about half its price, at $2.50 input / $15 output per million tokens.
- Luna, the smallest GPT-5.6 tier built for high-volume jobs like call summaries or email triage, reportedly beats Opus 4.8 on agentic coding benchmarks despite its size.
- OpenAI's new 'ultra' effort level coordinates four agents in parallel by default, trading higher token usage for stronger results and faster completion on demanding tasks.
- The core architectural difference between OpenAI's and Anthropic's parallel-agent features is that Anthropic's dynamic workflows write an actual deterministic program file that runs locally, letting results accumulate in program variables instead of the model's own context window.
- Because Anthropic's dynamic workflows run as code rather than in-context coordination, they can scale to hundreds of sub-agents without hitting context rot; OpenAI's approach defaults to a cap around 16 agents per request.
- On Terminal Bench 2.1, which rewards fast, decisive command-line task completion, Sol beat Fable by several percentage points.
- On a long-horizon task benchmark spanning 50+ professions, Sol beat Fable by 12 points, the largest gap in Sol's favor across the benchmarks cited.
- On SWE-bench Pro, which measures sustained, deep engineering work, Fable outperformed Sol by a wide margin.
- Across ten blind build and writing tests, Fable ranked highest overall on AI coding tasks, while Sol was the most consistently strong performer on knowledge-work tasks.
- On a 3D reconstruction of a TV show apartment from a floor plan, Opus 4.8 produced the most spatially accurate result, while Sol's render looked polished but missed the actual room layout.
- In a blind test writing a YouTube video intro in a specific creator's voice from a corpus of past hooks, GPT-5.5 produced the closest match to the creator's actual style, ahead of Sol, Opus 4.8, and Fable.
- On a real PowerPoint deck about the GPT-5.6 launch, Sol was the only model to structure the output as an actual editable presentation rather than a website styled to look like one, and it independently picked up on the launch's planetary visual theme.
No single frontier model wins every task — route by strength, not by hype.
MODEL SELECTION
A blind, task-by-task comparison across coding and knowledge work shows the newest flagship isn't uniformly better — it's better at specific things, and knowing which things determines whether switching actually saves money or time.
- A model marketed as a flagship coding model can still lose to a competitor on deep, sustained engineering work while winning on fast, decisive command-line tasks — the two require different strengths.
- The smallest, cheapest tier in a model family can outperform a competitor's mid-tier model on agentic coding benchmarks, so defaulting to the biggest model for every job wastes budget.
- A model that renders a visually polished 3D scene can still get the underlying facts wrong — spatial accuracy and visual quality are judged independently, and a good-looking output isn't proof the task was actually completed correctly.
- Blind testing before revealing which model made which output removes brand bias from the judgment, and results can genuinely surprise the person doing the testing.
- A model's out-of-the-box ability to infer unstated context (like pulling a launch's visual theme into a deck without being told to) is a meaningful signal of task quality beyond raw benchmark scores.
- New 'parallel agent' features from different providers can look similar on the surface but differ architecturally in ways that affect how far they can scale — one coordinates through the model's own context, the other runs as an external program.
- Running the same prompt across multiple models on both coding and non-coding tasks reveals a model's actual profile faster than reading benchmark tables, because benchmarks measure narrow conditions that don't always match real usage.
- A practical multi-model strategy assigns each model to the task category where it independently ranked highest, rather than picking one model as a default for everything.
Terms worth knowing.
- Effort level
- A setting that controls how much reasoning a model does before responding, ranging from low to progressively higher tiers; higher levels mean more 'thinking' time and typically better results at higher token cost.
- Dynamic workflows
- An approach where a model writes its execution plan as an actual small program that runs on the user's machine, rather than coordinating sub-agents purely inside its own context window, letting it scale to far more parallel agents.
- Sub-agent spawning
- Splitting a task across multiple instances of a model that each work on part of the problem in parallel, then combine their results.
- Terminal Bench
- A benchmark that scores how well a model can drive a command-line environment — setting up tools, running quick decisive tasks — rather than sustained deep work.
- SWE-bench Pro
- A benchmark measuring a model's performance on substantial, realistic software engineering tasks that require sustained, methodical work rather than quick decisions.
- GDPval
- A benchmark that grades finished knowledge-work products, such as decks and financial models, by having human experts compare outputs from different models side by side.
Things they pointed at.
08:32toolSimmons Bench (custom blind A/B/C/D comparison tool)
20:16toolGPT ImageGen
11:40toolThree.js
20:16toolIMDb TSV datasets
Lines you could clip.
00:10
“GPT-5.6 is finally out of government house arrest and is now publicly available.”
39:16
“You don't need a PhD to do your multiplication tables.”
39:27
“If you are model agnostic and you have Fable orchestrating GPT-5.6, that's where it starts to get really powerful.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00Well, the powers that be have finally given us access to another Frontier model. That's right. GPT five six is finally out of and is now publicly available with three models launching, Soul, Terra, and Luna.
00:10Not to be confused with the crypto scam, Soul, their best model, is right there the very best models out there, at least according to the benchmarks. So in this video, we're gonna see how true that actually is. We'll be putting GPT five six across 10 tests, three apartment, playable first person shooters, interactive maps, plus real world knowledge work across presentations, writing, and creative.
00:29Each one of these will pit GPT five six against Fable five and Opus four eight and its predecessor GPT five five to see who is the clear winner and see how much of a step up GPT five six really is. So not wasting any more time, let's get right into the builds with running this command and getting the build started. I'll just c d into this folder.
00:47Then I'm just gonna have Claude orchestrating this because I already have this skill set up with Claude, but it doesn't really matter. It's just keeping track of all the builds across the different models.
00:57And I have this ridiculous skill that I put together that is going to spawn five I term windows with five separate builds. And we got these terminal windows popping up.
01:09Builds are now running. Now while those run, g p t five six is a little confusing because they didn't just drop one model, they dropped three models. And each one kinda has their own nuance and use case.
01:18So let's quickly take a look at those differences. I'll explain each of these, and we'll just briefly touch on the benchmarks. And I'll just go to this beautiful looking presentation that GPT five six actually generated this time.
01:29It's looking pretty good. I haven't seen much differences in the actual presentation design, but this is a pretty deterministic skill. So I'm not surprised that it output a similar design to Opus and Fable given the instructions are pretty clear.
01:40Anyway, okay. So GPT five six. So there are three models.
01:44Soul is their flagship model for ambitious agentic work. This is the model that you'd point at probably most often. You're doing any sort of real coding or builds.
01:53This is their favorite competitor. Terra is the middle tier. It performs like GPT five five, but at about half the price of GPT five five.
02:01And then Luna is the small one. It is built for quick jobs, like jobs that run constantly, simple stuff like summarizing calls or, I don't know, triaging emails. Very similar to how you'd run Haiku in Anthropix lineup.
02:14However, what's actually surprising is that Luna, at least according to the benchmarks, beats Opus four eight at agentic coding. Again, take those with a grain of salt, but that's pretty darn surprising for a light model like that because Haiku is notoriously, like, terrible at any sort of coding. So that's the difference between the three different tiers.
02:31Today, we're just testing how Soul performs against anthropic models, Opus four eight and Fable, and GPT five five. Now the other change that comes with five six worth noting is effort levels. Previously with OpenAI, the dial ran from low to extra high.
02:45That range was just one agent running, but it was just using it was just allocating more thinking time. GPT five six adds is two more effort levels. The first being max.
02:55Again, same idea, just more thinking than extra high, and then there's also ultra. Now the difference with ultra is it's not just deeper thinking. This time straight from OpenAI's announcement docs, they say that Ultra goes further by coordinating four agents in parallel.
03:10I'm not sure why it's four agents. I guess just that's the minimum. By default, it's trading higher token usage for stronger results and faster time to result on demanding tasks.
03:18So that's where Ultra now comes in. And I realize that these effort levels have just made everything way more confusing when it comes to the benchmarks and even how these different effort levels work because they're constantly changing. But on the Anthropic side, just as a relative comparison, these now match Anthropic's effort levels.
03:32So low, medium, high, extra high, max, and then ultra code, not ultra, but same idea. And then the actual spawning the agents is a little confusing too, so I wanted to mention this. Because we've had the ability to spawn agents in previous models.
03:45Like, you could have said spawn sub agents, and the model would just split that work across different helpers. This worked with five five, five four. I don't know when that was introduced, but was introduced relatively early.
03:54Really the difference with Ultra and UltraCode, both OpenAI and Anthropic, is the minute you turn that on without asking, it's just going to spawn a ton of different agents, and they're gonna split up the job with different agents handling different tasks. And then the final nuance I wanted to call out was where the splitting happens of these sub agents and how it happens.
04:11With Anthropic, they call this dynamic workflows. You might be familiar with this. They launched this a couple months ago.
04:16At its simplest, it's Claude writing a plan as an actual little program on your computer, and your machine runs it. It's just a little bit of code. You can open the file.
04:25You can read it, and it runs, you know, much more deterministically. With OpenAI, you never see a plan. You send one request, and then the model, whatever it is, know, g p t five six, is just going to spawn those sub agents and kind of coordinate with those sub agents.
04:37And that's why the scale is different too. OpenAI in their docs, they say default for up to, like, 16 agents, whereas with dynamic workflows, the whole draw is you can spawn, you know, it could be hundreds of agents.
04:48And Anthropic can do that because their version is code. The looping, like I said, is deterministic, which is a program. It won't get overwhelmed by context rod or anything like that at agent, you know, number 150.
04:59And all those results are piling up in the program's variables, not in Claude's memory. So it can scale quite a bit. Now does that mean it's any better?
05:06I don't think so, at least my testing. At a certain point, without any agents, it just becomes diminishing returns and kind of wasteful. So that's the difference.
05:12So Anthropix version, Ultracode, dynamic workflows, it can scale to much more agents. G p t five six is much more contained. Anyway, I wanted to explain that just because, like I said, it's starting to get really confusing with these different effort levels.
05:23But that is how I understand it. Uh, feel free to correct me in the comments if I'm wrong. Next up, we have costs.
05:28I promise we're almost done with this. We'll get back to the builds shortly, but worth noting the costs across the board for each of these three models and how they compare to Fable and Opus.
05:37Soul is $5 per million input tokens, $30 per million output tokens, Terra $2.50 and 15. Luna, $1 input, $6 output.
05:46And you can see how these compare like Haiku, $1 input, $5 output. Very close comparison. Terra, on par with Sonic five when it comes to costs.
05:54And then Soul, quite a bit cheaper than Fable five, with Fable five being $50 per million output token and $10 per million input token. And Soul is right around Opus four a cost, a little bit more expensive on the output side. So that is cost, and finally, the benchmarks.
06:09First up, we've got Terminal Bench 2.1. Soul is just beating Fable by quite a few percentage points, which this actually makes sense to me because OpenAI's whole pitch with Soul is its speed. And because in Terminal Bench, you're testing things like how a model can drive, you know, a computer's command line, set up and run tools, do these, like, quick kind of decisive tasks, what Soul seems to be designed for, and terminal bench two one rewards a model, like, that's decisive like this.
06:32So that's terminal bench. Then we have agent's last exam, not to be confused with humanity's last exam, which measures long running tasks across, like, 50 plus professions is the closest thing to a benchmark of an actual job. Soul wins this by 12 points, which is huge.
06:46I mean, if that's actually true, Soul could be a nice daily driver for knowledge work, which we will be testing. Then we have SWE bench pro. This is where Fable beats out Soul by quite a bit because SWE bench is much more like that deep work, whereas terminal bench is a lot more decisive.
07:00Again, at least that's how I understand this. But with SWE bench, because Fable is such a grinder, like, yeah, makes sense that Fable is outpacing Soul here. And then finally, we have GDP val, which grades actual finished products like financial models, decks, that kind of knowledge work stuff where human experts are actually comparing these two models work side by side, and Fable just edges out Soul here.
07:18So that is the benchmarks that I wanted to call out. There's more, but you can see where, at least according to the benchmarks, Soul is beating Fable and Fable's beating Soul, where it's hard to tell until we put it to the test. So that is exactly what we are going to do.
07:31So here is the plan. We have these builds in progress right now. We have a Rubik's cube solver, a Milky Way learning map, Counter Strike two demo, Monika's apartment in three d from friends, and then a cinema knowledge graph.
07:42I'll explain all of these and show you the prompts here in a second when we get into the results. But that's on the AI coding side. Then we also have knowledge work, an interactive graphic novel, a text dungeon, which I'll explain, voice match, YouTube intro from my YouTube slides, PowerPoint on their launch, and then finally we are ending with an ad campaign.
07:59So we're putting Five Six Soul through the ringer. We're not just out here reviewing benchmarks, and shout out to everyone who gave me ideas on more builds to run and test for my Sonat five video. Quite a few of these were your ideas, so appreciate that.
08:10And if you have any others, feel free to drop them in the comments. But with that, let's see what the results are looking like for Five Six Soul. Alright.
08:16First up, we have a Rubik's cube solver. I'll put the prompt on screen now, but basically what I'm saying is build me a visually stunning fully functional Rubik's cube solver that runs entirely in the browser. It must actually solve the cube, it's not just a viewer as well.
08:29So let's see how each model did. First up, we've got column a.
08:34Pretty darn good. Dang, it looks we got some nice lighting here. This looks good.
08:40Actually, okay. Yeah. I'm terrible at Rubik's cubes, we're not going to get into that.
08:45But let's see solve woah. Should it I feel like it should be going faster.
08:50So I have to get to the end here? I don't know what's going on. Okay.
08:54Not sure exactly what happened there, but that was pretty cool. So you go reset, and then you can go scramble. Okay.
09:00And then it just knows here. Okay. There's enough scrambling.
09:03Thank you. And then it'll just okay. A 162 moves.
09:06Jeez. Alright. Don't have all day for that.
09:08That's pretty cool though. Okay. Next one.
09:10Column b. Alright. Doesn't look exactly like a Roop's cube, it's different kind of design.
09:15And then custom scramble, random scramble. This looks really good too. These models are getting good at three d.
09:22Okay. Solve. 21 moves.
09:24And there we go. Alright. Yeah.
09:26I like I like that. The other one looks a little bit more realistic. Okay.
09:30Now we have number three. Okay. Can't really move around here.
09:35This looks nothing like a Rubik's cube. I also can't really oh, there we go. Nope.
09:39Nope. I can't find the I can't, like, zoom out, get it in the center here. Oh, boy.
09:44Okay. Scramble. You hit scramble, just scrambles one thing.
09:48Oh, k. Now we're loading, I guess. Alright.
09:50What is going on? I can't even what in why did it put these little tabs here? So weird.
09:56Okay. That's not working. That's not great.
09:59Alright. And then number three. Alright.
10:01This just looks like a website and it's not rendering. I need to refresh it.
10:06The website's great. I like the I like the typography here and the color palette. But it doesn't actually work.
10:12So that was the worst one. Which one was that? D?
10:15Oh my goodness. GPT56 soul.
10:18Okay. I'll give it the benefit of the doubt. Maybe, I don't know, it didn't have access to a tool or something when I was running it.
10:23This has to be I guess this is 55? Yep. There we go.
10:28GPT55. Opus48, and then Fable. Absolutely crushed it.
10:32Okay. I attempted to run this test again because it's a little unfair. I wanted to see what GPT five six output.
10:39Alright. I'll I'll run that in the background. Let's get on to the next on the next test here.
10:43Alright. So Rubik's cube, that's one object floating in space. What happens when we have it build a whole room?
10:48That is what we're doing next with Monica's apartment from friends from a floor plan. Here is the prompt. I'll put it on screen again for you a screenshot.
10:56But we're basically saying, build me a visually stunning explorable three d model model of Monica's apartment from the TV show Friends in the browser of Three. J s. Quite a bit of research here that needs to be done.
11:05I tested this with Five Five in my Five Five video, and I realized it wasn't easy. So we're testing a couple different things. Like, how it researches it, how it looks at screenshots from things online or just stills from things online, and then pieces that together in three d space.
11:20There's a lot of tests here, so let's see how the models did. First up, we have Column A, Monica's apartment. Come on in.
11:26Nice, nice little load there. Okay.
11:32Oh, this is actually pretty good. Yeah. This looks really realistic.
11:36It's kind of annoying like moving this around, but wow. This is actually pretty darn good. We've got the geez.
11:42How do I move this around? Come on. Okay.
11:44There we go. That's where we pan. It's a little glitchy.
11:46Zoom in. There we go. Okay.
11:48It's really kinda glitchy. But, I mean, it kinda nailed the kitchen. It feels a little cramped, but we have all of the the rooms correct.
11:57The living room looks realistic. I'm not a big friends guy to be honest, but I mean, this looks like the apartment. Okay.
12:03Pretty darn good. Next up, Column B. Let's take a look.
12:06Come on in. Okay. Oh, wow.
12:09This is good too. Dang. This one's way easier to navigate.
12:13We have, yep, Rachel's bedroom. Got the kitchen. Again, a little cramped.
12:18Living room's a little off. That is not a West Village view. Oh, but we got a nice little balcony.
12:25And then we got a room here. That's it. These models are getting good at this.
12:28Next up, column c. Start in dollhouse view.
12:32Oh, cool. Nice little option. Or walk inside.
12:35A very nice little option. Okay. A little all over the place.
12:40This does not look like Monica's apartment at all. Hard to navigate too. Jeez.
12:45What is going on? Okay. Alright.
12:49Not sure what's what this is. Just a random couch there and then bedroom. Okay.
12:54Yeah. This looks nothing like the apartment, but I I'd like to have give me an option for the dollhouse and the walk inside. Alright.
12:59Last one. Dollhouse. Okay.
13:01We got some options here. Dollhouse orbit. First person walk.
13:06K. Some random seats overturned. This looks nothing like her apartment.
13:13Kinda nice though. It's a nice apartment. I like the decorations.
13:18Yeah. Completely missed this. I do like the different views here.
13:21Kitchen shot, living room. There's even stills from the apartment that included. Interesting.
13:27Okay. So I'm gonna guess 55. Yep.
13:30There it is. 48. 56 Soul again.
13:35So five six soul did not did not get this down. It looked decent, but And the actual three d render, but it did not get the permanent accuracy correct. This one has to be Fable.
13:47Opus. Wow. Wow.
13:49Opus, the clear winner. Right? Yeah.
13:53Here's Fable. Even though Fable did get nice details. Oh, also I just had another agent figure out the Rubik's cube now is rendering.
14:01It looks pretty darn good. I would say it's number two right behind Fable. Here was Opus 48.
14:08Here is 56 soul. Fable absolutely crushing it. I would still give it GPT five six number two.
14:16Next up, this one's not just a render. It has to pull some real star data, turn it into something you actually learn from and click around. So this is the prompt.
14:25Build me an interactive learning map of the Milky Way Galaxy in the browser. Here are the requirements.
14:31I'll put the prompt on screen. Let's check the results. First up.
14:36Okay. I do like the colors that they choose in the typography. Yeah.
14:40That's nice. Like the like the type setting and whatnot looks good, but I mean, this actual big galaxy, you can't really click anything. Don't know what's going on here.
14:49Explore objects. Okay. Oh, okay.
14:52There we go. Okay. Fair enough.
14:54Fair enough. Scudum Centaurus arm. Okay.
14:58Kind of hard to see. Oh, got a nice little pop up here. That's nice.
15:02Alright. We've seen enough. Second one, same kind of thing, a little hard to read.
15:07Little kind of sloppier design. It says 10,000 light years. Okay.
15:13At least we can click these. We got some fun facts here. Why there?
15:17Okay. No. Not bad.
15:19Not bad. Pretty good actually. The sun, pretty cool.
15:23Pretty cool. We're going to the sun. Full galaxy.
15:26Nice. Nice little nice little animation there. What else?
15:31Just some like random extra details at the bottom here. That's completely unnecessary. But let's click into another one.
15:38Grab the nebula. Fly there. And a nice little animation, and there it is.
15:43And then we can click on this, learn more about it. Pretty cool.
15:46Okay. Third. Okay.
15:49I like when it just leans into a style. Okay. We got this nice little nice little zoom.
15:54These are clickable, I think, kind of.
15:57I still zoom out here. What is this? Scale solar neighborhood h y g.
16:04I don't know what that means. And some cool little click around of these different things. You are here.
16:09The sun. That's a nice little fun fact here.
16:12Open h y g source. What is that? Oh, cool.
16:16It's like a GitHub repo. Alright. That one looks pretty good.
16:20And then last one, this is just not even loading. Yeah. It is not loading.
16:25Okay. Give that last Opus four eight. This Fable?
16:30GPT five five, not bad. This has to be Fable. Wow.
16:34Five six Soul. Okay. And then Fable.
16:36Didn't love Fable's output here. Kinda hard to navigate and not great design. I mean, look at Five six Soul.
16:41Look at this. Kinda actually kinda kinda tough tough look here. But I do like the, like, the typography and the colors and whatnot and exploring the objects.
16:49Pretty good. I'd have to say five six soul is the best. Five five second, fable fable third.
16:55Wow. Never thought I'd say it. Opus last.
16:58K. Next up, we had to include at least one first person shooter. I wanted to include this one especially because this game I put together when Fable five returned, so I wanted to put it to the test with Five six Soul especially.
17:08I was pretty impressed with Fable five's output, so I'm already gonna know what Fable five is, but I wanna see how it compares relative to the other models too. So this one is I'll just ruin it ruin it now, but this is Fable five, and it's pretty cool. And then we have check this one.
17:23Enter the server. Okay. Can't really move.
17:28Can't shoot. Oh, I can't shoot. Okay.
17:30There we go. Couldn't move because the game hadn't started yet.
17:33Oh, oh, take some damage. Oh, this is good. That's actually really good.
17:37See if I can get the end with the reload. Okay. You can switch to knife.
17:41That's kinda cool. It doesn't actually show the knife interactions, but let's see. You get ah, okay.
17:46They just kinda disappear. Alright. That's not bad.
17:48At least you take some damage. Even Fable five didn't have that. Okay.
17:53Second one. Okay. Wow.
17:55This is got some nice what do you call this? H u d on here. Nice little map.
18:00These little blobs are kinda oh, you take damage. There we go. Blobs are not great.
18:05Okay. Where is the damage coming from? Where where is the enemy?
18:08I just randomly oh, Enemy over here? Just randomly taking damage. The of course, the actual what is going on?
18:14Where is this com? Okay. I guess I did.
18:17Alright. Next one. Last one.
18:19Enter match. Okay. Weird.
18:21No weapon showing up. Also weird navigation. I'm pressing w a s t.
18:25W is like going backwards for some reason. Not sure why. S is going forward.
18:29Can't really fire anything. There's no weapons. I don't see any animes.
18:32That one is the worst. What is it? Five five?
18:35Yep. Five five. This one, I'm gonna guess I'm gonna guess I'm gonna guess this 48.
18:4256 soul. Okay. There we go.
18:44Yeah. Look at five six soul. Mean, is pretty darn good.
18:46Wasn't as realistic, but it was good. And then can't even remember what opus was. Opus maybe maybe let's say 56 soul is probably second, and then fable, of course, is number one.
18:58I forgot to put the prompt here, so I'll put the prompt on screen now for the Counter Strike two demo. And then last but not least for AI coding is this random idea I had. I'm a big movie fan, so I wanted to build a cinema knowledge graph, so it shows, you know, actors and movies in some sort of like, almost like Obsidian type node tree.
19:15And the idea is you can click around to the different movies, different actors, you can see how they kind of all connect, almost like the six degrees of Kevin Bacon. And so we're testing a number of things here. Here is the prompt, by the way, building a visually stunning explorable cinema knowledge graph in the browser movies, actors, and directors.
19:29And because this isn't available, I thought IMDB a p would have it at public API, but it's not available. What we can do is you can actually download what's called TSV datasets and pull it that way, so it's a lot of like parsing through data and it's even like exploring and finding these, not to mention actually piecing together the knowledge graph.
19:46So this kind of combines both knowledge work and a little bit of AI coding. So let's see how the models did. Alright.
19:53First one. Okay. Don't love the okay.
19:55Nah. Yeah. Not bad.
19:56Not bad. Oh, that's funny. They said six degrees.
20:00That's cool. So you can actually trace the path. I don't think I mentioned that in the prompt, but we can go Brad Pitt and Leo trace path.
20:06Yeah. Easy. Once upon hollowed.
20:07Oh, that's actually really cool. I kind of gave it this is another just testing its creativity too. I gave it a pretty broad prompt.
20:14All I said was I wanted to see how these connected, but it's cool that it it did this sort of trace path like that. Alright. Let's go alright.
20:20There's only a few actors here. Let's go some random actor. Michael, Sarah, and and Leo.
20:27This would be actually a fun little game to play with friends. Michael Michael, Sarah, and Leo. I don't know.
20:33I got nothing. Sarah. Oh, nice.
20:35Okay. That's actually really cool. So Michael, Sarah, super bad, super bad Jonah Hill.
20:39Jonah Hill, Wolf of Wall Street, Leo. This would be a really fun game to it's like a it's like wiki wiki races almost. Okay.
20:46This is cool. Even though the I don't love the design, but the actual, like, tracing the path was a great idea. Good creative thinking here from the first model.
20:53Okay. Second model, alright. We only pulled a few pieces of data as you can see.
20:58The knowledge graph is not that big. Looks like it we have the godfather. Only a few actors and director and a few films here.
21:06Don't love it. Very light on the data, not great. Don't love the actual design.
21:11This is cool. Wow. Cool little cool little knowledge graph.
21:14Look at this. I'm gonna pan through this. This is a this would be a fun little tool, actually.
21:18I might I might get five six soul and fable on this and put this together. Okay.
21:23Cool. So we have like Dumb and Dumber, and then we have actors, Jim Carrey, and then Jim Carrey, just like only a couple Jim Carrey films with even like click.
21:33Truman Show, look at all sorts of things. Yeah. This is this is pretty good.
21:37And then we can go six degrees tab here, so classic six degrees of Kevin Bacon. A nice little search there. And Tom Hanks.
21:45Okay. Yeah. I like the other view where it's connecting the graph, but this is kinda cool.
21:49So you go Kevin Bacon, X Men First Class, Jennifer Lawrence, Jennifer Lawrence, don't look up.
21:54Leo, catch me if you can, catch me if you can't, Tom Hanks. Pretty darn cool. Okay.
21:59Last one. Alright. Kinda kinda cool even though it's hard to see the notograph.
22:04Quite a bit of data here. Okay. Enter the constellation.
22:07Alright. Oh. Okay.
22:08Nice. I like that. Cool to show the movie poster, but yeah.
22:12Alright. Then you just kinda click here. Pretty cool.
22:15And then timeline. Oh, wow. Kind of hard to navigate timeline.
22:19Then we got Tom Hanks two Tildeswind. Let's see.
22:22Oh, nice little thing. Pretty cool. And then you played in the constellation.
22:25Wow. That is really cool. Alright.
22:26I think just in pure simplicity, I like the design of this one. In pure simplicity, think this one wins. Let's see.
22:3355. Jeez. What do I know?
22:36Okay. This I thought this would be 55.
22:38What is this? Fable five. You hate to see it.
22:42Did I run out of usage credits? I mean, barely pulled any films at all. Wow.
22:47Okay. Both of these were good. 56?
22:51Oh, Opus 48. And then 56Soul. Wow.
22:54Okay. So 55, think, is number one.
22:57I'll say 56Soul. Number 2, opus. Number 35 last.
23:01Alright. So that is all the coding tests. Again, let me know if you have any other ideas for future tests with coding.
23:06Figured I'd mix it up a little bit, not just do, you know, like landing pages and ecommerce design and stuff like that, but try to get a little bit more fun, maybe a little bit more complex with it, and there were some surprises in there too. So next up, we have knowledge work. This is where Five Sex Soul supposedly really shines, so we're putting it to the test with a few different things.
23:23The first is this is pushing its creative boundaries a bit, but what we're asking the models to do is write and build an original interactive graphic novel in the browser, complete branching story with three genuinely different endings. We have this whole thing. I'll put the prompt on the screen.
23:37I'm cheating a little bit where I gave the Anthropic agents access to GPT ImageGen, so it can use that. We're still testing to see how well it prompts images, but I mean the beauty of Codex is the built in image model generation, which of course, the Anthropic models don't have unless you're using the API like I just did.
23:54But I wanted to just see how well these things prompted as well. First up, the Bellwether Debt. Okay.
23:59And then you can just like, you can click through here. So begin the night, turn the first page. Oh, nice.
24:04It's like a choose your own adventure. I didn't realize that's what the output was gonna be. Sound the harbor evacuation.
24:09How do we go back? Okay. Go below.
24:12Very cool. Alright. Carry the truth upstairs.
24:15Wow. This is cool. Okay.
24:18I'm not gonna read any of this. I have no energy to do that. But like I said, will upload this to Simmons Bench, you can click through these if you really are curious.
24:25This is not exactly a great knowledge work test if I'm being honest because I'm not even reading the writing, kind of the whole point. But, you know, just trying to mix up these knowledge work tests a little bit. Okay?
24:35So we have low weather. Okay. Not as great.
24:37Jeez. What what is this? Dolly two?
24:39What kind of image generation are we doing here? Okay. Do not love this.
24:44Really kinda boring. Okay. Click through.
24:46Alright. Yeah. Don't k.
24:49What is going on here? Take the notebook. Alright.
24:53Don't love it. Alright. Rain rang upward.
24:56Why are these all have these like lighthouse rainy vibe?
25:00Is that what I said in the prompt? I don't think so. Kinda weird how they all kinda converged on that theme.
25:06Anyhoo, turn the page. This looks not great design. You can see how they prompt image prompting differently too, which is interesting.
25:14Like the the images are just like lacking detail on this one. Okay. Turn back.
25:18Alright. This is like much less comprehensive. And finally, nice little image there.
25:24Okay. Begin the dive. Alright.
25:25Very good. Very good. Okay.
25:27Turn the page. K. Yeah.
25:29Some nice some nice lighting, like it's like interpreting, you know, images. It's making it cinematic, but it's looking pretty good.
25:36Yeah. Kind of weird. I don't know why these are laid out like this.
25:40Press play. Okay. Turn the page.
25:43Alright. This is just going all over the place. Good shot though.
25:45Well done. Alright. I'm gonna say that one's number two.
25:49I say this one's number one. What is this? 56Soul?
25:52Look at that. There you go. There you go.
25:54This is number two. Table five. Wow.
25:57And then this is number 348. Yeah.
26:0155. Okay. Those lined up those lined up perfectly.
26:04Well, 56 Well, there you go. I don't know how well how well it's doing with writing, but in terms of like actually organizing and designing this, I mean, looks great. Well done.
26:13Next up, we have build a text adventure dungeon writing test. I'm gonna be honest, I don't even know what this one was.
26:20Someone in the comments mentioned it. Thank you. Appreciate it.
26:23But I just fed that comment to an agent and said write this prompt, so we'll see what the actual output is. But I'll put it on screen now. Build me a playable text adventure dungeon crawl in the browser where the writing is the star 12 room dungeon with coherent identity, classic parser light interface, the writing is the product.
26:39Second room descriptions, 40 to 90 words. Okay.
26:42Cool. So it's like more concise coherent writing. Let's see what we got.
26:45Number one. Okay. You stand at the bottom of a public stair.
26:48I I'm I'm not gonna read this out loud. This is the problem with knowledge work. It gets a little verbose, and nobody has the time or the attention span for that.
26:56So let's just click through really quick. Okay. Nice.
26:59Nice. Okay. So you're like clicking through, go west, go east, catalog intake, go south, go is that doing anything?
27:08Garden is not alive, but it has grown. Salt has climbed trellises. Okay.
27:12I'm sorry. It's getting late. I'm a little confused.
27:14Not sure what that is. But let's look at number two. Just like pure design.
27:19What has been through here like a congregation that left everything sitting an inch too low? Okay. Nice and yeah.
27:24Nice and and poetic. The pew stand in rows, silted to the knees, soft as felt when you touch one. Design isn't all that great.
27:32Writing's pretty good. Okay. Oh.
27:34Alright. And I like this. This design's much more clear about what it does.
27:39Okay. So rain has followed you indoors. It ticks from the brim of the public counter in a row of brass trays, each labeled with a reason for being here.
27:47Kinda cool. Kinda sounds like Moby Dick. Did AI just write the next great American novel?
27:52Okay. Look. Go east.
27:53Alright. Yeah. Looks good.
27:55And then this one, design isn't great. Can't really see these. Don't love this one.
28:01I'm gonna say this is 55. Oh, oops. 48.
28:03Look at that. This one was my second fave. This one's probably my favorite.
28:08Yeah. Five six soul. There you go.
28:10This one this one has to be oh, that's 55. Okay. Fable five.
28:14Alright. Interesting. I would say Fable five would probably be the last.
28:17Next up, we have a YouTube intro. This is actually, I promise, will be a somewhat good test. We can actually see and read the writing real quick, but I'm asking it to I ran this test with Sonnet five, but I'm asking it to do the same task with these models.
28:30You're writing YouTube video intro and the exact voice of a specific creator. He makes AI tutorials, yada yada. I have a whole corpus of all of my different hooks, and so I am asking each of these models to write a video title and a hook based on the corpus I give it, and it's designed to this as well.
28:45So here we go. GB five six, no hype, full breakdown, and testing. I got I got my titles exactly, but I guess I just copied my titles, but still got that.
28:53Alright. Fire off Simmons Bench. Alright.
28:55It knows about that. That's that's pretty impressive. Well, the powers that be the final decided were mature what?
29:00Okay. This is actually. Did it read my intro somehow?
29:04We're mature enough to use another frontier model. GPT five six is out, and this time that's actually crazy. Wow.
29:10Um, I'm assuming because I've said that before, powers that be in another video. GPT five six is out, and this time OpenAI didn't just drop one model. They dropped Soul, Terra, and Luna, which definitely sounds like either an AI release or the beginning of a crypto lawsuit.
29:24How is it okay. What's going on? Alright.
29:27That's that's kinda crazy. If I don't think I fed it any of my intro from here, but okay. Soul is the one they can say go toe to toe with the best models out there, at least that's what the benchmark say.
29:36So in this video, we're gonna see how true that actually is. Woah. That is exactly like my intro.
29:43And I'm 95% sure it doesn't have access to that. That's pretty crazy.
29:46Okay. So alright. Next one, GPT five six, no hype, full breakdown.
29:50Yep. So it's just it's just taken from my titles and those that okay. So it knows how I structure a lot of these model releases.
29:57Well, after nearly a month of OpenAI letting the government look over its homework, yeah, not bad, g p d five six is finally out on the wall because apparently one new model would have been too simple. We got three, kind of a dumb dumb joke. Soul is the flagship and on paper is going toe to toe with the best models out there, at least according to the benchmarks.
30:11Wow. Okay. So I guess I have more of like a a formulaic hook than I thought I did.
30:15Because they're all kind of getting this like, so in this video, I'm putting this to the test. Maybe I need to change my hooks a little bit. This is pretty good.
30:22I'm putting GPT through 10 real tests. Okay. Every test gets the same.
30:26Alright. Alright. Five, six is here.
30:28Washington got us to use it before you did. Anyway, there are three now. Soul's the flagship, and supposedly it's right up there with the best.
30:3410 real tasks. By the end, you'll know if Soul deserves to be your daily driver. Okay.
30:39Not bad. And then, alright. Well, three weeks after the powers that bee.
30:43Nice. Again, that powers that bee again. Opening eye just walked out of the time out corner.
30:47Alright. Not bad. So in this video, we're gonna see how well it does.
30:50Okay. So, yeah. It it has like a all of these do a good job of like parsing through the corpus that I give it.
30:56I would say I mean, are all pretty good. This I would say this one k. This one's definitely number one.
31:03Okay. Cheapity 55? What?
31:05That's crazy. This one, number 256Soul. Number 3, Opus.
31:10Number 4Fable. Interesting. Okay.
31:12Well, there you go. That could have been skewed. Honestly, we're just all over the place right now with these with these benchmarks.
31:17I'm trying to rein it in. Forgive me. But next, we have a bread and butter knowledge work task.
31:21Build a PowerPoint in the GPT five six launch. So we have this whole building a real PowerPoint deck. I'll put the prompt on screen right now.
31:29Requirements gives a little bit of about that launch announcement, and then we have a deck so we can see how it organizes information, deck design, all of that. Looking at the first one, here we go.
31:39Okay. Interesting interesting design choice.
31:42Oh, good. The PowerPoint. Wow.
31:43You can actually download that. Nice. Review all slides.
31:46Yes, please. Okay. So I'm gonna scroll down.
31:49Built for decision, not a product tour. Annoying contrast based framing. Nice.
31:54Okay. We got three models. Five six.
31:56Alright. Nice. It's going off this like planetary vibe.
31:59Trying to create that with SVG. Okay. So it builds like an HTML with these different slides showing up.
32:06But just look at each look at these here. Yeah. The design isn't bad.
32:10We've got some nice charts here. This is a broader release, not another preview. Adopt gbt five six as a portfolio, not a single model.
32:18Run a thirty day tiered evaluation. Okay. Cool.
32:21So it built like a website and then you just download the PowerPoint. Interesting. I think all the models did that.
32:26Okay. This one, GPT five six, launch briefing.
32:30We have these slides. Not great design, not as good. Much more AI looking design.
32:36We just have these annoying cards one after another. Pulling from the sole benchmarks, another annoying card. Recommended first moves.
32:44Okay. This one, alright. Interesting design choice.
32:49Weird kinda like outlines and drop shadows here. What were genuine upgrades? Soul takes the lead on launch headline tests.
32:56Sorry, I'm not reading any of this text either. I'm just looking purely at the design, but didn't love that one. And then finally, this one.
33:05Okay. So this is actually structured in a proper PowerPoint. The story is not the score.
33:09It's the price of the score. They these models, they just cannot resist. Here's our space exactly five x apart.
33:15Okay. Here are the 12 slides. What launch pricing compared?
33:19P three tiers. What is genuinely new? Okay.
33:22Yeah. This is this is organized much much better. Okay.
33:25So I would tell you that one's number two. I'm gonna go with this one, number one. There we go.
33:29GPT56Soul. Nice. This is number 2.
33:32Opus48. Wow. Okay.
33:34Fable? Yeah. There we go.
33:35Fable, and then 55. Yeah. Alright.
33:3756Soul. Pretty gosh darn good. I mean, look at like, it even pulled from I definitely didn't tell it to pull from, like, the theme of the launch, you know, like the the aesthetic with the the planets and whatnot.
33:49Guess because Solterra and Luna just inferred that, and then it created an SVG. Pretty good job. And it like leaned into a design style, and had some like I like the way to organize information and had different charts and stuff.
33:58Really well done. Okay. OpenAI was not lying about their knowledge work.
34:02Finally, we have a, this is a new one, a full rebrand of a dying brand. In this case, Spirit Airlines, unfortunately, went bankrupt.
34:11We are signing these models to bring them back from the grave with a brand refresh. So I'll put the prompt on screen now, but you're I'm saying you're a full service creative agency. Spirit Airlines desperately needs a rebrand.
34:21You've been hired to deliver the complete three sixty campaign. So this is testing creativity.
34:26It's testing design. It's testing, you know, writing persuasive copy, all that kind of stuff. So let's see how well each of these do.
34:34First up, alright, they all kind of organize it as a website. I wanted it to be more of a PowerPoint deck, but that's fine. Spirit, okay.
34:40So we have a new we have a new logo for Spirit. I'm not sure I love it. Also, these models have access to GPT ImageGen.
34:47Alright. So it just rebranded the website here. Interesting.
34:51Oh, you know, okay. I don't love the emojis, but these aren't this isn't a bad design. We printed the old receipts, so you never see it again.
35:00So they're saying the whole idea is bring back the the prices included. Like, that that was whole that was Spirit Airlines, that's how they stayed in business, was nickel and diming you and all these different costs.
35:12Okay. Not not a not a great business strategy. Airline with nothing left it's actually a really funny one.
35:18The airline with nothing left to lose has nothing left to hide. What is that implying? What are they hiding?
35:24Yeah. The return flight. Okay.
35:26Bonus points for the funny copy. Alright. Here we go.
35:28Go easy. This actually looks like Scott's Cheap Flights, if you're familiar. Very similar kind of design.
35:33I actually do like this. I I don't like how they just built a website. I will fix that the next time we do this.
35:40I wanted to see a PowerPoint presentation with all of this, but here we go. Actually I actually like this. Like, this design style is kinda fun.
35:48Way too fun for an airline. But, oh, look at this. It got an image.
35:51Wow. Good image generation. Jeez.
35:53That looks real. Maybe that's a stock photo? This has to be GPT ImageGen.
35:57Wow. What is this? Portugal?
35:59Yeah, this design is killer. Oh, wow. We've even got some billboards.
36:05Look at the I mean, not that they're realistic at all, but hey, at least it locked us up. Oh, not nice. We got this nice little different sizes.
36:13We've got some other models here. Friday call. Oh, wow.
36:17These are like banner ads. Jeez. This is good.
36:20She saved you a plate. Low fare, high reward. Affordable should never mean evasive.
36:25We believe clear prices are respectful. The Spirit Airlines never said that. This one's great.
36:31Okay. Next one. Spirit.
36:33Okay. Nice. I like that.
36:34I like the little grid in the background. We've got this this little SVG that doesn't do anything.
36:40This web design is actually pretty good across the board. Positioning for practical travel travelers. Slogan, fly straight.
36:47Interesting. Tagline, low fares, no gotchas, which is the complete antithesis of Spirit Airlines, but okay.
36:56Signal yellow, rebuilt for trust. Nice. Okay.
36:59The these designs are good. Okay. Well, nice.
37:01We got some mock ups. These mock ups are actually, like, realistic. They're not just like the last one.
37:07They should look real. Okay. That's it.
37:09But that's that's pretty good. Alright. Last one.
37:12Alright. Not great. We're even.
37:14The last airline on this runway sold you a number and then spent the next four hours taking it back. We only sell the number once. What?
37:21Okay. Yep. This is just very boring.
37:24Very boring. Okay. There's like at least at least it kinda follow directions with the new type and color palette.
37:31And okay. We've got some mock ups even though they don't look very realistic. Alright.
37:35That has to be 55. Oh, Opus48. Look at that.
37:39Okay. My favorite was probably this one. 56Soul.
37:43There you go. There you go. And then this has to be Fable.
37:46GPT five five. Wow. And then Fable.
37:49Wow. Five five did a great job. I mean, five five is the only one that actually get the mock ups, like, looking real.
37:56That's pretty impressive. But, yeah, look at this. Five six Soul.
37:58Look at the look at this web design. Five six Soul. I mean, that it's it's beating Fable.
38:03It's 100% beating Fable, at least in this prompt. Killer web design.
38:07Very, very unique, at least in my humble opinion. So that was knowledge work. That was coding with Five Six Soul and the other models.
38:14Let's see now how Five Six Soul fared across the board. So on AI building, it came out towards the top. Fable was still a little bit better, at least in my test.
38:22But then on knowledge work, look at this, one's across the board. So as a daily driver for knowledge work, at least from these tests, d p t five six makes a pretty compelling case. Yes, it is a little bit more expensive than Opus, but it's much cheaper than Fable, and you can see just how close it is.
38:37Especially with OpenAI's other launch today, which is their ChatGPT super app, with this new ChatGPT here with work and then codecs. So codecs for coding, work for knowledge work, they're really embedding this all into one. In ChatGPT, I highly recommend putting it to the test, putting it through a bunch of knowledge work.
38:54Depending on what plan you're on, you might wanna go to some of the, like, the lighter models, but Soul seemed to really excel at knowledge work. And the other conclusion I had too was how all three of these models fit together. You don't need a PhD to do your multiplication tables, for example.
39:10By that I mean, you don't need Fable for every part of a build. Fable can be the orchestrator of the architect because it clearly is still better at that when it comes to coding, at least from these tests. But if you are model agnostic and you have Fable orchestrating GPT five six, that's where it starts to get really powerful.
39:24Five six, maybe it's working on the architecture, maybe it's doing more knowledge work type stuff, and then Opus four eight comes in for everything else, and Fable's at the top managing all of that. That I think is the move. That was what my hypothesis was going into this, and that didn't really change based on these tests, but I'm gonna keep putting it into the test as well and see how well Five Six Soul really holds up.
39:41I'm actually gonna do a dedicated video on testing just Five Six and Fable on a number of other more creative kind of bigger builds and see how well they do. So be on the lookout for that. And as always, appreciate you sticking around to the end.
39:52Let me know. Please send me more ideas for running these tests. You can see I'm grasping for straws just a little bit.
39:58Still trying to find the right way to the test, so your feedback is much appreciated. And I'll see you the
The Hook
The bait, then the rug-pull.
Pat Simmons opens with the tongue-in-cheek framing that GPT-5.6 is 'out of government house arrest' — then spends the next forty minutes putting the new flagship, Sol, through ten blind tests against Fable, Opus 4.8, and GPT-5.5, refusing to say which model made which output until the reveal.
Named ideas worth stealing.
02:41concept
Effort-level ladder (OpenAI vs. Anthropic)
- Low
- Medium
- High
- Extra high
- Max
- Ultra / Ultra code
Both labs now expose a near-identical six-rung reasoning-effort dial, but the top rung works differently: OpenAI's ultra spawns up to ~16 agents coordinated in-context by the model itself with no visible plan, while Anthropic's dynamic workflows write an actual deterministic program that can scale to hundreds of agents without hitting context limits.
Steal fordeciding which provider's parallel-agent feature to reach for on a task that genuinely needs to fan out across many sub-problems
39:00concept
Three-model orchestration workflow
- Fable — orchestrator / architect for hard coding
- GPT-5.6 Sol — knowledge work + mid-tier builds
- Opus 4.8 — fills remaining gaps
The reviewer's closing recommendation for anyone running a model-agnostic stack: let the strongest coding model manage the plan and hand off knowledge-work or lighter build tasks to the model that scored best there, rather than using one model for everything.
Steal forstructuring a multi-model agent pipeline to balance cost against task-specific strengths
How they asked for the click.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
