ChatGPT

Our verdict on ChatGPT

ChatGPT is the tool we reach for to write things that other tools will use: PRDs, build prompts, song prompts, scoring prompts. It is also the API behind several of our internal apps. When the output has to be code or a visual, we have picked Claude almost every time, and in January 2026 Jordan said we were moving more of our spend to Anthropic.

Our view has moved around. In October 2025 Jordan said OpenAI had "taken the lead on the chat bot for sure" after GPT-5. Three months later the focus on ads and enterprise pushed him the other way.

What we built with ChatGPT

• Ryz Labs Video Studio (June 2025): Jordan used ChatGPT to improve his first Replit prompt, then built an "enhance with AI" button that rewrites prompts for whichever video model you pick. The first version shipped in a few days.
• Ryz Labs Autoblogger (July 2025): ChatGPT and Claude write the posts and OpenAI's image models illustrate them. About 150 posts per brand so far.
• Ryz Score (July 2025): Supabase edge functions send a scoring matrix, the job description, the resume and the full application to ChatGPT, and the score goes back into our ATS.
• NTRVSTA (August 2025): ChatGPT and other AI tools handle scoring, translation and feedback for the AI interviewer.
• Public API trading bot (August 2025): every 30 seconds the bot sends price, holdings and risk settings to ChatGPT, which decides whether to buy or sell. Built in three to six hours; $300 became $299.
• Sam's ATS (August 2025): a generic PRD from ChatGPT got the first version out of Claude Code.
• n8n sales outreach workflow (October 2025): ChatGPT turns research into a structured, custom email to each CEO.
• LLM Math Roaster (November 2025): GPT-5 was one of four models writing Lean proofs, and ChatGPT judged them all. It gave Gemini 98 and itself 70.
• AI start-of-care packet generator (December 2025): Sam got a PRD from ChatGPT, pasted it into Google AI Studio and had a prototype in about thirty minutes.

How we use ChatGPT

1. Write the PRD in ChatGPT, build somewhere else. Sam's ATS and the start-of-care prototype both started this way (Episode 9, at 03:15).
2. Improve your first prompt. Jordan tells ChatGPT "I'm making a new app in Replit" and asks it to improve the first prompt (Episode 1, at 22:21).
3. Tell it which model the prompt is for. The Video Studio's enhance button names the target model. For Suno, ChatGPT writes the style prompt and lyrics; Jordan asks it not to name artists, because Suno won't accept them (Episode 3, at 13:41).
4. Turn one thread into a mini app. Sam keeps a calorie and macro tracker in a ChatGPT project: a photo plus a voice note per meal, and it learns what he usually eats.
5. Give the API structure and keep guardrails outside it. Ryz Score sends a fixed scoring matrix. The trading bot sends portfolio, history and buying power, while stop loss, max position and a cash float live in the app's settings.
6. Turn feedback into a task list. Jordan's tip: screenshot each page, add the recruiter's feedback, ask ChatGPT for a feedback list, then hand it to Claude Code (Episode 9, at 09:34).
7. Keep a human check on anything it publishes. Jordan warned that ChatGPT can hallucinate or miss what a brand is, which can hurt search traffic (Episode 2, at 12:01).

I mean, that's interesting because it makes, like, you know, One ChatGPT conversation is like a mini application. Right? And now it's not a conversation to you anymore.

— Jordan Metzner, Episode 4

Where ChatGPT falls short

• Visual output. Given the same 57-page 10-K, Claude built a dashboard right away. ChatGPT's summary was "too much to read" and took more prompts to get four graphs (Episode 11, at 12:41). For his cholesterol visualizer Jordan picked Claude for its web rendering (Episode 13).
• Coding. A week after GPT-5, Jordan said it was not a better coding model than Anthropic's. For OpenAI code we use OpenAI Codex, not the chat app.
• The GPT-5 router. Sam kept getting the quick answer when he wanted the thinking one. Jake Trefethen and Emily Kurtz of Public had the same complaint.
• Context it doesn't have. Jake pointed out the trading bot didn't know it only got two round-trip trades a day, so it wasn't trading with that in mind.
• Judging. Carina of Axiom noticed one model had only proved the theorem for natural numbers, and suggested compiling the Lean proofs instead of trusting an LLM judge.
• Images. Jordan was underwhelmed by Images 2.0 next to Nano Banana; Sam liked its text fidelity (Episode 40).
• Ads. OpenAI exploring ads made Jordan want to spend more time with Anthropic, and turned Sagi Waitzman off OpenAI (Episode 28).

I think I mean, it's disappointing. It doesn't it's not a better coding algorithm, and it didn't beat, you know, the Anthropic model, so seems like everyone's gotta get back to work.

— Jordan Metzner, Episode 8

Guests are mostly positive. Hooman Radfar called ChatGPT a "brand new step function change." Tristan Wilson sees OpenAI enterprise accounts becoming common in construction. Dr. Chadi Nabhan thinks patients who research symptoms in it are better informed.

ChatGPT compared

• ChatGPT vs Claude: Claude won our dashboard, visualizer and coding comparisons.
• ChatGPT vs Gemini: Gemini topped the Math Roaster, and Bito's Amar Goel called Gemini 3 good but less rich as a product.
• ChatGPT vs Nano Banana: Jordan saw no clear difference between Images 2.0 and Nano Banana for his episode backgrounds.

Episodes featuring ChatGPT

• Episode 1: prompt enhancement in the Video Studio.
• Episode 4: ChatGPT as tool of the week, from calorie tracking to a restaurant's digital menu, plus Ryz Score.
• Episode 8: the trading bot and GPT-5 reactions.
• Episode 11: the 10-K test against Claude.
• Episode 22: ChatGPT as the judge in the LLM Math Roaster.
• Episode 29: why we are shifting spend toward Anthropic.
• Episode 53 (October 2026): Jordan has tried GPT-6's lower-level models through Codex and found them cost effective.

FAQ

Is ChatGPT good for building apps?

We use it to plan apps, not to code them. Sam wrote the PRD for his ATS in ChatGPT and handed it to Claude Code, and Jordan uses it to improve the first prompt he gives Replit.

How do you use ChatGPT as a calorie tracker?

Sam keeps one thread inside a ChatGPT project, sends a photo and a voice note every time he eats, and gets calories, macros and suggestions for the rest of the day. He found it simpler than MyFitnessPal and didn't pay anything extra for it.

Is ChatGPT better than Claude?

Not for our coding and visual work. Claude built a better 10-K dashboard in Episode 11, and Jordan said GPT-5 was not a better coding model than Anthropic's. In January 2026 he said Ryz Labs was moving spend from ChatGPT toward Anthropic.

Can ChatGPT run a trading bot?

Jordan's bot sent his portfolio and the IBIT price to the ChatGPT API every 30 seconds and let it decide to buy or sell. It made about ten to twelve trades and turned $300 into $299.

What is ChatGPT bad at?

In our episodes: long summaries with weak visuals, a GPT-5 model router that gave quick answers when Sam wanted deep thinking, and a risk of off-brand or made-up content if you auto-publish without review.