How I AI: My GPT-5.6 Sol Benchmark & 4 Game-Changing Workflows (vs. Fable)

I have been very, very sad the last week because I did not have access to my true favorite, top-of-the-line model, GPT-5.6. But guess what, babes, it is back! I am here to walk you through GPT-5.6 Sol, Terra, and Luna. I’m going to tell you what these models are, how I’ve been using them, and why Sol is my heart's favorite.

There was a period where we didn't have access, and I found myself desperate to get this workhorse model back. But this isn't just about my opinion. We are going to run the very famous, very new How I AI vibe review benchmark against common tasks, from PRD writing to prototyping to whether or not it’s cute in my OpenClaw agent. I'll tell you very scientifically if this is the model you should be working with all the time.

For weeks, I've been testing the entire GPT-5.6 family—Sol, the next-gen frontier model; Terra, the balanced model for everyday work; and Luna, the affordable one for high-volume tasks—against its biggest competitor, Anthropic's Fable 5. This episode is my love letter to GPT-5.6 Sol. I blind taste-tested these models, and Sol consistently came out on top for the kind of work I do every day: building real, functional, and interesting products. It's not just theoretically smart; it's practically effective.

The 'How I AI' Vibe Review Benchmark

I got kind of bored of the old vibey vibe checks, so I built an extremely scientific How I AI benchmark to test models on the tasks I actually perform. It covers:

• PRD Generation: How well does it write a product requirements document?
• Prototyping & Wireframing: Can it develop fully designed, robust prototypes for various apps?
• Coding & Debugging: Can it handle an agentic, multi-step debugging process?
• Agentic Voice (Chit-Chat): Can it talk to me like a human?

I tested the three versions of GPT-5.6, Fable 5, and Sonnet 5. The eval harness runs these tasks against each model and uses GPT-5.5 as a tough LLM-based judge. But it also gives me a page for the 'Claire Taste Test,' where I review all the assets, look at the designs, and score them myself.

The Claire-Weighted Index

This is my show, so I get to strike the balance. I decided on a 70/30 split: 70% my taste, 30% the machine's judgment. And with that split, your girl loves GPT-5.6 Sol. It had the highest taste score by a significant amount. I went through dozens of evals, and I just have to say, I really like this model.

Fable 5 isn't out of the game, though. When I don't have to talk to it (I hate talking to Fable 5, it talks to me like an engineer that has never met a human before), it outputs pretty good work. Tara and Luna did fine. Sonnet 5 landed at the bottom, though I’ve found a couple of specific use cases where it shines.

Per-Task Winners

Here’s how it broke down by category:

• Prototypes: GPT-5.6 Sol, hands down. The designs were more functional, more interesting, and less 'slop'. Across a doc scheduler, a dev tools site, and a consumer app, Sol's outputs had better visual hierarchy and interactivity.
• PRD Writing: GPT-5.6 Terra was my favorite. It’s streamlined and to the point. If you want clean, crisp, direct business writing, Terra might be the way to go.
• Bug Hunting: The LLM judge thought Sonnet 5 did the most complete and accurate job here, though I'm still refining this part of the benchmark.
• Agentic Voice: Sonnet 5 gets a gold star. Aside from its love of em-dashes, it sounds the most human. I still struggle to get the GPT models working well with my OpenClaw setup.

One of the most revealing parts was comparing the full-fidelity prototypes. Here's a look at a dense operations dashboard for a doc scheduler app. Sol is on the left, Fable on the right.

Fable's version is a pretty standard dark-mode, monospace layout. But Sol's is unique—a clean, neutral color layout with great visual hierarchy and semantic color. More importantly, it was functional. Everything I expected to click and work actually worked. This was a consistent theme.

Workflow 1: Building a Gamified Homework App in One Shot

Now for some fun. My kids are coin-operated. My middle child is basically going to be an enterprise sales rep. If he does his homework, I need to give him a Skittle or let him trade Skittles for Nerf guns. My husband even sent me an XP system proposal via OpenClaw. So, I took that OpenClaw-generated PRD, dropped it into Codex, and let GPT-5.6 Sol generate a fully gamified homework tracking system.

The result was pretty ambitious and surprisingly polished for a one-shot build.

• Student Dashboard: It created separate summer quests for each of my kids. My oldest earns progress toward a one-on-one basketball coach for practicing piano, while my youngest works toward Minecraft rewards. It includes a focus mode with a timer to track their time on tasks like Math Academy.
• Gamification: It went all-in. They can earn XP, unlock companion avatars like 'Beat Bot' and 'Comet Fox', and gain 'power auras'. It even gamified collaboration, allowing them to earn more XP by working together.
• Parent HQ: This is where it really impressed me. It generated a full parent dashboard where I can review their progress, turn quests on and off, edit point values, and add new tasks and rewards. It listened to my inputs—basketball for my oldest, Minecraft for my youngest—and built the reward system around them.

Is it a perfect consumer-grade app? No. But it's a lot better and more complete than what I've seen from other models in a single attempt.

Workflow 2: Breaking Through Fable’s Pedantry

If you take one thing away from my comparison, it's this: Fable is theoretically hyper-intelligent, and Sol is practically effective. As a manager, I struggle with brilliant colleagues who can't see the forest for the trees and never ship. That's what working with Fable can feel like.

I was working on an integrated prototyping tool inside ChatPRD and ran into a wall. Fable had built a very hardened, rigid tool-calling loop. It was so rigid that only GPT-5.5 would run; I couldn't get Sonnet, Opus, or any other model to work.

I ran eval after eval. Fable was insistent:

"No, bro, that's, it's, it's totally these models, model's fault."

I was convinced it was our problem. As soon as I switched to Codex with GPT-5.6 Sol, I gave it a simple prompt:

"Look, I'm just not convinced we can't get Sonnet 5 to work. This is ridiculous. Just do what you think is correct."

In one shot, Sol fixed it. It got out of its own head and got the other models working. That unlock—the willingness to reconsider its own limitations to achieve a user goal—is exactly what you need when you're building products. Fable's obsession with technical precision broke itself, while Sol's practicality got the job done.

Workflow 3: Rapid Video Editing for Social Media

This is one of my new favorite time-savers. I do a lot of social clipping, and it's incredibly tedious. Recently, I spoke at a Cursor event on the future of PM and got the full video recording back.

To create a hype video, all I had to do was drag the file into the Codex interface and ask:

Can you cut this video into five clips for social?

I gave some follow-up feedback, asking for the clips to be horizontal, faster, and tighter. In minutes, I had sharp, funny clips ready to go. I just dropped them into CapCut, added some music, and shipped them. This would have taken me so much time to do manually. The ability to just drop a video file in and have it edited is a huge productivity boost.

Workflow 4: Automating My Life with Browser Control

Finally, the last and best use case. GPT-5.6 Sol is a BEAST when it comes to browser use. I am deeply obsessed with letting Codex plus GPT-5.6 control Chrome. If you haven't tried this, you just type @Chrome in Codex on a logged-in page and give it instructions.

I'm sorry, LinkedIn, but I had to try this. I opened my LinkedIn page and told it:

"Can you use Chrome to reply to messages that are a very high value to ChatPRD or the How I AI podcast? Keep the bar very high... only accept them if they're executives of tier one companies."

It burned through probably 500 messages. It replied to people I needed to get back to and sent thank-yous to people who said nice things about the podcast. It just rocked through it. I’ve also used it to test web apps and fill out annoying forms. When I got rolled back to GPT-5.5 for a few days, my life was measurably worse. Please, please, please learn to use @Chrome and just let GPT-5.6 rip.

Sol is the Model for Getting Things Done

So that's the very scientific How I AI model benchmark. GPT-5.6 Sol is my favorite model right now for its practical effectiveness. It's an excellent writer, it builds fantastic and functional web apps, it can unlock complex technical work, and its browser and video skills are top-notch.

While Fable might be technically brilliant, Sol is the partner I want when I need to ship. A quick tip: you will notice Sol loves a forest green. I think it’s a system prompt called “woodland elegance” or something. I love green, but if you don’t, you might need to prompt against it. It's a clear 'tell' for this model.

I'd love to hear what you think about these models and what I should add to the benchmark. We'll be publishing all the benchmark outputs to the ChatPRD blog.

Production Thanks

Production and marketing by Penname. For inquiries about sponsoring the podcast, email jordan@penname.co.

Resources Mentioned

• Find me: Website, LinkedIn, X
• Tools: GPT 5.6 (Sol, Terra, Luna), Codex, ChatPRD, CapCut, Math Academy
• References: Cursor event talk, ChatPRD Blog