GPT 5.6 Soul vs Fable 5: A Real-World Head-to-Head, Not Another Benchmark

Benchmarks tell you how a model scores on a leaderboard. They don't tell you what happens when you hand two models the same prompt, walk away, and come back to see which one actually shipped something that runs. So that's exactly what we did.

In this test, GPT's 5.6 Soul and Claude's Fable 5 went head-to-head on the same challenge: build a fully functional Lovable clone, running entirely locally, with no API keys. Everything had to run off our existing ChatGPT and Claude subscriptions through OAuth, not retail API credits. And as a bonus, both models took a swing at cloning Sonic the Hedgehog.

Listen, we build in public here, so this is the whole story, including the parts that didn't go to plan.

The Setup

Two fresh terminals side by side. Codex (running GPT 5.6 Soul) on the left, Claude Code (running Fable 5) on the right. Both got the exact same prompt: create a Lovable clone, fully functional locally, free to deploy a development Convex database, working end to end.

An open source cost meter tracked both sessions in real time. That meter, by the way, is its own build-in-public story. It started as three separate projects (a GPT meter, a Claude meter, and a Grok meter) that got merged into one Model Meter repo, and the merge left a couple of bugs that surfaced live during recording. One was a display quirk where anything under $10 got three to four decimal places. Another was a hardcoded assumption that the ChatGPT account was on the $20 a month plan instead of the $200 a month Pro plan. Both got fixed on camera, mid-demo. The plan detection fix was the fun one: rather than being told the answer, the AI figured out the correct plan by reading the plan-type claim inside the OAuth ID.

Where the Two Models Split Immediately

The first meaningful difference showed up before either build finished.

GPT 5.6 Soul assumed the Lovable clone's AI generation feature would be powered by an API key. That breaks the entire premise of the test, and it took a follow-up prompt to correct: no API keys, use the subscription. To its credit, GPT confirmed a REST API wrapper around the CLI was possible and built it.

Fable 5 caught the conflict on its own. It detected that requiring an API key would violate the local-only constraint and proactively planned a workaround that shells out to the locally authenticated Claude CLI. As said in the video: "Fable detected that being an issue and it asked me. ChatGPT assumed that I would just input an API key to make it functional."

Early upper hand to Fable for planning ahead.

The Results Flipped the Script

Here's where it gets interesting.

GPT 5.6 Soul finished in about 23 minutes at a cost of $6.19, and it spent a noticeable chunk of that time running Playwright browser tests. A test Chrome window kept popping up as it verified its own work. That testing phase is almost certainly why its cost jumped, and it's also why the result worked. The clone, served on localhost 5173, loaded, took prompts, and built.

Fable 5 took its sweet time, wrote a whole bunch of code, and named its creation "Ember, a local AI app builder with a warm forge aesthetic," complete with a fresh Convex dev deployment and a self-reported list of "three gotchas I hit and fixed." And then the preview didn't load. At all. A full build with no working output.

"Fable, what did you do wrong? Was there just no QA?"

That's the inversion that makes this test worth writing about. Fable 5 is normally the more thorough, more expensive model. This time it came out cheaper than 5.6 Soul and failed the one check that matters: does it run. Meanwhile Soul paid extra for verification and delivered a working app. The lesson is blunt: money spent is not the same as verified output, and a model that tests its own work end to end is worth the extra tokens.

Stress Testing the Working Clone

To see if GPT's clone was actually useful, it got a real task: build a landing page plus lead magnet for an HVAC company based in Las Vegas, Nevada.

It delivered a clean, working site. Simple, yes, and it would need imagery and fine-tuning before going in front of a client, but functional. A follow-up prompt asked for a gamified call to action where visitors enter their home's square footage and get an AC cost estimate, and the clone incorporated it successfully. The estimates themselves would need reference data to be trustworthy, but the mechanics worked.

"It's not perfect, but that's impressive. And it just works."

The Bonus Round: Sonic the Hedgehog

The Sonic clone comparison, run before recording, showed the more typical pattern between these two models.

GPT 5.6 Soul built its version fast and cheap: $2.24. It was playable, but it barely resembled Sonic. It looked like Sonic without feeling like Sonic.

Fable 5 ran overnight and cost $15.90 in retail-API-equivalent credits. But the result nailed the gameplay physics, the visual character, and the details that matter, like coins scattering when you hit an enemy or obstacle. It even added an endgame boss that nobody asked for. "Cost more, but did a much better job."

That's cost correlating with depth, which is the usual relationship. The Lovable clone result was the anomaly.

What the Numbers Say Across All These Tests

Across the ongoing side-by-side tests behind this series, the pattern is consistent: nine times out of ten, Fable 5 costs more than 5.6 Soul, and usually by around 5x. Put differently, 5.6 Soul typically lands at about one fifth of Fable's cost for a comparable prompt. The Lovable clone, where Soul cost roughly twice what Fable did, was the exception, and the extra spend went directly into Playwright QA.

Speed is the other consistent finding: Anthropic models are noticeably slow compared to GPT and Grok. Maybe that's just being spoiled by the competition's speed, but it shows up in every comparison.

The Takeaway: Match the Model to the Job

This is not a "use X, not Y" verdict. Fable 5 is still in daily use here and excels in plenty of areas. The honest conclusion from watching both models work is about matching the tool to the task:

• Everyday builds that don't need maximum reasoning: 5.6 Soul, all day. It's faster and usually about one fifth the cost.
• Complex, polish-heavy, or overnight deep work: Fable 5 earns its premium, as the Sonic clone showed.
• Either way, watch for QA signals. A model that runs browser tests before declaring itself done beats a model that claims completion with no working preview, regardless of price.
• Skip retail API costs for local testing. Both builds ran entirely on subscription OAuth by shelling out to the authenticated CLI. That's a real pattern worth stealing for your own local tooling.
• Overnight agent runs are a superpower. Give a model a detailed task at night, approve permissions up front, and start the morning by previewing what it built.

The reality is that the average person isn't using these models at full capability. So run your own side-by-side tests with identical prompts, track the actual dollars, and judge the output by whether it runs, not by a leaderboard.

The model meter used throughout this test is dropping as an open source repository on the Clearmud GitHub. If you want it released sooner, say so in the comments on the video.

Watch the full head-to-head, including the live bug fixes and both Sonic clones, here: https://www.youtube.com/watch?v=YnrGNJClLyE