Latest / AI Fire Daily / #65 Robin: Fable 5 vs GPT 5.6 Sol - The "Manager vs. Worker" Breakdown for AI Devs
Transcript
- 0:00What happens when you pit the world's most creative
- 0:02AI against the most efficient one? Oh, it is
- 0:05the ultimate battle. It really is. It is the
- 0:07battle of judgment versus execution. Yeah. Welcome
- 0:11to this deep dive. Today, the entire landscape
- 0:14just shifted. Completely shifted. OpenAI just
- 0:18dropped GPT 5 .6 all. Right. And they released
- 0:22it alongside two other models. Terra and Luna.
- 0:26Yeah, the planetary launch page was incredible.
- 0:28It was a beautiful page. It had these tiny fives
- 0:30and sixes drifting around. It looked incredibly
- 0:33polished before I even touched the actual model.
- 0:36Well, the benchmark charts also looked completely
- 0:38pristine. They really did. Sol beat Opus and
- 0:41Fable across the board. And it came in at a drastically
- 0:44cheaper price point. Which changes everything.
- 0:47Exactly. It makes every comparison feel very
- 0:50different now. For developers and creators, it
- 0:52raises a very real question. Which model actually
- 0:55helps you build better things? Right. And how
- 0:57do you do it without burning your budget? We
- 0:59aren't just looking at cherry -picked benchmark
- 1:01screenshots today. No. We are unpacking a grueling
- 1:05full -day test. A side -by -side test. We compared
- 1:09GPT 5 .6 Sol directly against Fable 5. Yeah,
- 1:14we ran Sol and inside Codex. And Codex is OpenAI's
- 1:17dedicated coding environment. Right. And we ran
- 1:19Fable 5 directly inside Claude Kind. That is
- 1:22Anthropic's terminal tool for building software.
- 1:25Right. We tested them across four real -world
- 1:28coding and design tasks. We wanted to see how
- 1:30they handle actual production work. We needed
- 1:34to know if Sol can truly replace Fable 5. Is
- 1:36Fable still the model you opened for serious
- 1:39quality? That is the big question. We started
- 1:42by testing their raw creative instincts. Because
- 1:45basic coding is getting pretty easy for these
- 1:47models. Oh, absolutely. But making good product
- 1:49decisions is still incredibly hard. Yeah, the
- 1:52first prompt was deceptively simple. Very revealing
- 1:55though. We told them to build an open world browser
- 1:57bike game. We wanted something genuinely fun
- 1:59and playable. Right. We gave both models total
- 2:02creative freedom. We provided no strict design
- 2:04direction at all. No camera angle. No art style.
- 2:08We just wanted a game you could actually play.
- 2:10This tests much more than just basic coding ability.
- 2:15Yeah. It checks how well the model makes product
- 2:17decisions. It has to figure out controls, physics,
- 2:20and game feel. So how do they approach the blank
- 2:23canvas? Well, GPT 5 .6 sole. built a top -down
- 2:28bike game. And it actually worked perfectly on
- 2:31the very first try. Yeah, which is always a really
- 2:34good sign for execution. It chose a top -down
- 2:36perspective for a specific reason. Its internal
- 2:39guardrails prioritize functional, error -free
- 2:42code. Top -down mechanics are mathematically
- 2:44much simpler to compile. Exactly. It wants to
- 2:46guarantee a working final product. It had working
- 2:49movement and proper steering mechanics. You could
- 2:51collect coins and do little bunny hops. It even
- 2:53had tricks and a speed boost feature. Yeah, and
- 2:56the city layout was actually fully playable.
- 2:58Buildings blocked the player character properly
- 3:00during gameplay. But, you know, the camera made
- 3:03it feel a bit awkward. A top -down view is just
- 3:05weird for a bike game. Very weird. The controls
- 3:08were not as smooth as you want. It felt like
- 3:10a limited, slightly rigid prototype. Right. It
- 3:13did exactly what the prompt asked for. But it
- 3:16lacked any real soul or fun factor. Soul gives
- 3:19you the mechanics of a bicycle. Fable gives you
- 3:22the feeling of the wind in your hair. Wow. Yeah,
- 3:24exactly. Fable 5 went in a totally different
- 3:26direction. It built a 3D point of view style
- 3:29game instead. It took a massive creative risk
- 3:31with spatial reasoning. It simulates a 3D environment
- 3:35because it understands modern gaming. Right.
- 3:37It maps the concept of fun to modern game tropes.
- 3:41It generated a much bigger city map. with massive
- 3:44ramps. The big difference was the overall feeling
- 3:46of the game. Fable's version felt significantly
- 3:49more open and expansive. Yeah, the camera made
- 3:52the game much easier to intuitively understand.
- 3:54It had that real open -world energy you actually
- 3:57crave. You genuinely wanted to ride around and
- 3:59explore it. Table 5 clearly made the better creative
- 4:02call here. Oh, totally. A browser game can technically
- 4:05work and still feel boring. Fable 5 avoided that
- 4:08trap completely by understanding human enjoyment.
- 4:12But we really need to look at the hard data.
- 4:14Right. Cell took about 23 minutes to build its
- 4:17game. And it used roughly 31 ,000 tokens. Tokens
- 4:20are just chocks of data the AI reads. Yeah. Fable
- 4:24took about 21 to 22 minutes to finish. But it
- 4:27used a massive 90 ,000 tokens. It processed way
- 4:30more data. To think about the fun factor. That
- 4:32huge computational difference showed up in the
- 4:35final price. It really did. Soul cost about $4
- 4:38.50. Fable cost over $14 for the exact same prompt.
- 4:43Is token efficiency always the enemy of creative
- 4:46ambition? Well, open -ended prompts require a
- 4:49lot more computational thought. The AI has to
- 4:52actively explore more ideas to find the fun.
- 4:55Right. It burns tokens, deciding what makes a
- 4:58game actually enjoyable. If you restrict the
- 5:00tokens, you restrict the creative exploration.
- 5:03So brilliant creative leaps just cost more compute
- 5:06power. Exactly. Solship's a working version for
- 5:08much cheaper. But Fable has a far stronger instinct
- 5:11for creative product feel. If Fable burns 90
- 5:15,000 tokens finding the fun factor, what happens
- 5:17next? Yeah, what happens when we ask for something
- 5:20purely aesthetic? Does that computational cost
- 5:22actually translate to better visual design? We
- 5:26moved to a test explicitly demanding visual impact.
- 5:29We needed to see if that price gap remains justified.
- 5:32The prompt was designed to be dangerous in a
- 5:34good way. We told them to build a scroll -stopping
- 5:36interactive website. We gave them no fixed layout
- 5:39and no required theme. We gave them room to make
- 5:42massive creative choices. We wanted something
- 5:45polished, memorable, and absolutely worth sharing.
- 5:48Both models actually chose a very similar overarching
- 5:51theme. Yeah, they built 3D scroll -driven stories
- 5:55about a cosmic journey. They both included space
- 5:58visuals and dramatic scene transitions. They
- 6:01even integrated sound into the browsing experience.
- 6:04The basic idea was incredibly strong on both
- 6:07sides. Right. This was not a case where one model
- 6:09completely failed. The difference came down to
- 6:11the final emotional feeling. Fable 5 delivered
- 6:14an immersive emotional wow moment. The scroll
- 6:17motion and 3D background worked together perfectly.
- 6:20Fable orchestrates complex animations with a
- 6:23director's eye for pacing. You immediately felt
- 6:25like it was actually cool. Yeah. Sol's version
- 6:27was also very clean and highly useful. It had
- 6:30an interactive background and a decent story
- 6:32structure. But it completely lacked that visceral
- 6:35visual punch. Sol just stacked basic visual elements
- 6:38together logically. Mm -hmm. Mm -hmm. It did
- 6:41not craft an emotional journey for the user.
- 6:44For a prompt built around being impressive, Fable
- 6:46won easily. It understood the emotional goal
- 6:48of the task much better. Sol just gave you something
- 6:51nice, standard, and functional. Let us talk about
- 6:54the underlying economics here. The economics
- 6:56are completely wild when you break them down.
- 6:59Fable took 23 minutes and cost $19 .24. Sol took
- 7:04just seven minutes and cost a single buck. Two
- 7:07secs silence. Whoa. Building a fully functional
- 7:11interactive 3D website for one single dollar
- 7:13in seven minutes, that completely rewrites the
- 7:16economics of prototyping. It absolutely does.
- 7:18It changes how teams will prototype forever.
- 7:21I have to push back on the value proposition,
- 7:23though. OK, sure. Fable made a much better website
- 7:25overall. Yeah, it did. But was it almost 20 times
- 7:28better to justify the $19 price tag? Well, you
- 7:32have to consider the contest of the user. OK.
- 7:34If you are an agency pitching a Fortune 500 client,
- 7:38$19 to secure a multi -thousand dollar contract
- 7:41is absolutely nothing. That makes perfect sense
- 7:44for enterprise budgeting. Fable is designed for
- 7:46those high stakes, client -facing pitches. Sol
- 7:50is for your internal Tuesday morning update meeting.
- 7:54Where do you draw the line between needing a
- 7:56masterpiece versus needing a fast first draft?
- 7:59You always have to look at your primary goal.
- 8:01Use Fable when you absolutely need that first
- 8:04impression wow factor. Use Sol when you just
- 8:07need a cheap, usable starting point. Use Fable
- 8:09to dazzle clients. Use Sol to draft fast. Right.
- 8:12And that framing changes how you plan entire
- 8:15projects. Sols gives you something incredibly
- 8:17solid for a fraction of the cost. Definitely.
- 8:19We know Fable can build one beautiful thing.
- 8:21Yeah. We know Sol can build one cheap thing.
- 8:24Right. But what happens when we force them to
- 8:26multitask? What happens when they have to show
- 8:29actual range? We wanted to see them handle multiple
- 8:32concepts simultaneously. The prompt was to design
- 8:35five fundamentally different visual pieces. They
- 8:39could be games, simulations, or interactive visual
- 8:42experiments. I still wrestle with prompt drift
- 8:44myself when I ask models for variety. They always
- 8:47want to fall back into repeating the same core
- 8:50idea. That is exactly what happened with Fable
- 8:525. Really? Yeah, it struggled deeply to maintain
- 8:56focus across five distinct concepts. Its attention
- 8:59mechanism basically breaks down across varied
- 9:01contexts. It gets totally lost in the sauce of
- 9:04its own creative output. It built a very uneven
- 9:06gallery with highly confusing pieces. Fable had
- 9:09this chaotic art experiment energy. It made a
- 9:12buggy arcade ring game called Orbit. A hallucinated
- 9:15menu bug completely blocked you from playing
- 9:17it. Yeah. It made a space interaction toy called
- 9:20Singularity. But it had a very confusing goal
- 9:23for the player. Right. It also made a flight
- 9:25simulator called Terra. That flight simulator
- 9:28had very odd visual rendering. It also built
- 9:32a generative art tool called Inkflow. It just
- 9:34looked entirely random and largely unfinished.
- 9:37Finally, it made an underwater scroll story called
- 9:40The Descent. It was nice, but it felt incredibly
- 9:44familiar. Nothing in Fable's set really stood
- 9:46out as exceptional. Some pieces worked, but others
- 9:49felt heavily half -finished. Sol delivered a
- 9:52much cleaner gallery with true variety. It acted
- 9:55like an organized intern presenting a neat portfolio.
- 9:58Yeah. Sol built a typing game called Glyph Heist.
- 10:01It built a northern light music toy called Aurora
- 10:05Orchestra. It built an interesting life simulation
- 10:07called Tidepool. It built a map explorer called
- 10:10Atlas of Lost Echoes. And it built a very clean
- 10:12presentation slide deck. The deck honestly looked
- 10:15like something Claude would historically design.
- 10:17Saul definitely still had some software bugs
- 10:19in his code, though. Right. The typing game had
- 10:21no clear fail state programmed in. The music
- 10:23toy had controls that suddenly stopped working
- 10:26entirely. Tight pulls seemed to slow the machine
- 10:28down heavily. But the overall packaging was vastly
- 10:31superior to Fable. Every piece had a different
- 10:34shape and a clear purpose. The gallery interface
- 10:37made the work much easier to browse. I expected
- 10:40this multi -piece task to take the longest time.
- 10:43But it was actually the fastest run for both
- 10:45models. Sol finished the entire task in just
- 10:48seven minutes. It used about 22 ,000 output tokens.
- 10:52And it cost roughly one single dollar. Fable
- 10:55took 15 minutes and cost about... $15. So, actually
- 11:00wins this round completely outright. That was
- 11:02a huge surprise to see in the data. Why did a
- 11:05supposedly weaker model handle ambiguity and
- 11:07range better than the premium model? Well, Sol
- 11:10is highly obedient to structured formatting instructions.
- 11:13Ah, I see. It closely follows the prompt without
- 11:16wandering off creatively. It keeps things strictly
- 11:18on track when asked to diversify. Yeah. Fable
- 11:21just overthinks everything and loses the plot.
- 11:23Sol strictly follows the rules. Fable gets lost
- 11:26in the art. Exactly. When you need several different
- 11:29directions quickly, Sol is absolutely unbeatable.
- 11:32Beat. Let us pause here for a moment. Sure. Mid
- 11:34-roll sponsor, Reed. We are back to the deep
- 11:37dive. All right. Let's get back into it. We have
- 11:39looked at deep, multi -step, agentic builds.
- 11:43Those are systems that plan and do tasks on their
- 11:45own. But the reality of Enterprise AI looks totally
- 11:48different. Yeah, the reality of Enterprise AI
- 11:50is high -volume, single -shot questions. You
- 11:52send one simple API request to the server. You
- 11:55need one good immediate answer in return. There
- 11:58is no long loop or complex game to build. Who
- 12:02wins the mundane day -to -day work? We tested
- 12:05quick non -agentic API calls for this. These
- 12:08are just single requests asking for a direct
- 12:10response. The final results were an absolute
- 12:13blowout. Really? Saul won 24 times. Fable won
- 12:16only three times. Wow. But Saul did not win on
- 12:20raw intelligence. No, it did not. It won because
- 12:23Fable's heavy guardrails caused massive internal
- 12:26issues. Fable constantly refused to even answer
- 12:28the benign props. High -tier models are heavily
- 12:31fine -tuned to avoid giving bad advice. This
- 12:33aggressive safety training makes them overly
- 12:35paranoid. They frequently refuse completely safe
- 12:38everyday coding prompts. They think they're protecting
- 12:41you, but they're just breaking. When both models
- 12:43actually answered, the quality gap was tiny.
- 12:46Sol scored a 0 .98. Fable scored a 0 .966. That
- 12:50score difference is basically a statistical rounding
- 12:53error. The cost difference was absolutely massive,
- 12:55though. Sol cost about $16 for the whole test.
- 12:58Fabel cost about $63 for the same run. Sol answered
- 13:01more often and cost far less money. Its per -token
- 13:04pricing was roughly half of Fabel's rate. It
- 13:07makes you rethink OpenAI's entire naming strategy.
- 13:11They called it GPT 5 .6 Sol instead of GPT 6.
- 13:15That heavily hints it's just an execution step
- 13:18up. It is not a massive frontier leap just yet.
- 13:21No, it feels closer to a very strong, practical
- 13:24execution model. It might be better compared
- 13:26to Anthropix Opus 4 .8. Fable 5 is positioned
- 13:29as a true, elite frontier model. But a 10 out
- 13:33of 10 model is completely useless if it refuses
- 13:35to work. Right. If it refuses a 5 out of 10 difficulty
- 13:38task, it fails you. If you are a developer listening
- 13:41to this, you know this pain. Oh, absolutely.
- 13:43Reliability beats raw capability when you are
- 13:46scaling software. A genius model is useless if
- 13:50it refuses to answer your prompts. You do not
- 13:52always need a genius for basic server tasks.
- 13:56A cheaper model that handles the job reliably
- 13:58is significantly smarter. How should developers
- 14:01factor in refusal rates when budgeting for AI?
- 14:04Well, latency and refusals completely break automated
- 14:07enterprise workflows. They stop digital production
- 14:10lines totally cold. Right. That makes expensive,
- 14:13paranoid models a huge liability for high -volume
- 14:16tasks. Unreliable genius is useless for automated
- 14:20high -volume daily tasks. That is the clearest
- 14:23business lesson here. Sol had a much lower median
- 14:26response time. So it often felt significantly
- 14:28faster for these repetitive tasks. But the variance
- 14:31was higher with Sol's performance. Yeah, some
- 14:33runs could jump around a lot more in latency.
- 14:35Fable was much more consistent when it actually
- 14:37decided to answer. Fable usually landed around
- 14:39a tight 20 -second response window. But for simple
- 14:42API tasks, Sol is just much easier to justify.
- 14:45Looking at the bike game, the website, the gallery,
- 14:48and the API calls. A perfect mental model finally
- 14:51emerges for how to categorize these two AIs.
- 14:55Yeah, Fable 5 is the manager or the startup co
- 14:58-founder. It has refined taste, high -level strategy,
- 15:01and deep creativity. It will actively push back
- 15:03on your bad ideas. It is highly opinionated about
- 15:06design and function. It acts like a seasoned
- 15:08consultant, saving you from yourself. It is vastly
- 15:11better for brainstorming, creative builds, and
- 15:14product taste. It prevents you from over -engineering
- 15:17your own projects unnecessarily. GPT 5 .6 Sol
- 15:21is the excellent worker. It is fast, incredibly
- 15:24cheap, and highly token efficient. It is highly
- 15:27obedient for clear, repeatable, and mundane execution.
- 15:31It excels at careful verification and following
- 15:33strict formatting instructions. It stays much
- 15:36closer to the prompt and structured tasks. The
- 15:38overall capability tier becomes very obvious
- 15:41when you map it out. GPT 5 .5 is at the bottom
- 15:43of this new tier. OPUS 4 .8 is a strong, practical
- 15:47model slightly above that. Then you have GPT
- 15:505 .6, Sol as the better execution model. And
- 15:53Fable 5 sits way above them with a large capability
- 15:57jump. Fable has a much higher capability ceiling
- 15:59overall. Sol just has far better everyday economics
- 16:03for high volume execution work. The smartest
- 16:05workflow isn't choosing one model forever. No,
- 16:08not at all. It's using Fable 5 to carefully plan
- 16:10the complex architecture. Right. And then deploying
- 16:13an army of GPT 5 .6 Sol agents to execute the
- 16:18code. You use Fable to write the master spec
- 16:20sheet. You use Sol to actually write the tedious
- 16:23Python code. You use Fable when you need pure
- 16:26unadulterated judgment. You use Sol when you
- 16:29just need to ship the product fast. You let the
- 16:31manager draft the brilliant master plan. You
- 16:34let the worker hammer the digital nails. Benchmarks
- 16:37are helpful, but your own daily workflow matters
- 16:39more. You have to test both models inside your
- 16:42real software setup. You must weigh your project
- 16:44budget and your tolerance for speed. If AI is
- 16:47naturally stratifying into managers like Fable
- 16:50and workers like Sol, what does that make you?
- 16:52That is a great question. To succeed in the next
- 16:55five years, your job is no longer to be the worker.
- 16:58Your job is to be the executive director managing
- 17:00this hybrid AI workforce. You have to know when
- 17:04to deploy taste and when to deploy speed. It
- 17:06is all about intelligent resource allocation
- 17:08now. Exactly. Thank you for joining us on this
- 17:11deep dive. OTRO music.