Gemini 3.5 Flash vs GPT 5.5: I Was Surprised by Flash

Hear me out. Apples to apples, this was never going to be a fair fight. GPT 5.5 is the heavier model and would win most verticals on paper. But I wanted to see what happens when you let both models think for themselves with very minimal input. No PRD. Just a brief overview, a clear audience, and three artifacts to ship.

This is the writeup of that test, and yes, Flash surprised me.

The Setup

Two CLIs side by side. The anti-gravity CLI on the left running Gemini 3.5 Flash on high. The OpenAI codex CLI on the right running GPT 5.5 on low reasoning. Same prompt, same business, same expectations.

The business: Bay Area HVAC Pros. A residential HVAC company serving Oakland, Berkeley, and Alameda. Three artifacts under one brand:

• A customer-facing landing page to convert homeowners into booking service calls
• An internal dispatch dashboard for managing technicians and jobs
• An embeddable HVAC sizing calculator for the marketing site

The setup prompt told each model to pick a clean, trustworthy visual style, apply it consistently across all three, and propose its own directory structure and brand decisions before writing any code. I also pinned each project to a port (3030 and 3031) so I could pre-open two browser tabs and load previews the second they finished.

That was it. No logos, no typography spec, no color palette, no component library. Surface level on purpose.

Why No PRD

Some of you are already fuming. I get it. GPT 5.5 on low reasoning is not a real benchmark. Neither is skipping the spec.

But that was the whole point. The moment you over specify a prompt, you stop testing the model and start testing instruction following. With a PRD, every model converges on the same thing because you wrote the thing. Without a PRD, you actually see how each model reasons about a business, picks a visual identity, and structures a project.

Vibe friendly prompting is a legitimate evaluation mode. It exposes reasoning that a tight spec would mask.

Artifact One: The Landing Page

Flash finished first. Visibly. It like flashed across our eyes. GPT followed shortly after.

First impressions, I had to lean Gemini. The styling was not perfect, but it produced a clear call to action banner across the top with a 24 hour emergency perk and a contact path for anyone who did not want to scroll. It set the tone fast.

GPT's landing page was cleaner in places but read too plain. No banner urgency, no immediate hook for an HVAC customer in a hurry.

I was not expecting Flash to win here. That was the first surprise.

Artifact Two: The Dispatch Dashboard

When the complexity went up, the verdict flipped.

Gemini's dashboard was more ambitious. It wired the landing page form to populate the dashboard, so a booking on the marketing site appeared as a job in the queue. It supported drag and drop reassignment between technicians with statuses for on route, scheduled, on site, and unassigned. Gemini understood the mission, not just the screen.

GPT 5.5 went simpler and cleaner. The onboarding flow felt more polished and the layout was easier to read. I love the simplicity of GPT's here, even if it did less.

When I stress tested GPT's booking flow by typing nonsense phone numbers like 9999 9999, the validation did not fully wire up. So polish on the surface, gaps underneath.

Call it a split. Flash for the connective tissue across artifacts. GPT for the clean internal tool feel.

What This Actually Tells Us

A few takeaways that held up across both builds:

• Flash is genuinely fast. The speed gap is real and visible when you watch them side by side.
• Flash leans visual and marketing. It nailed the first impression on the landing page and grasped how the three artifacts should connect.
• GPT 5.5 leans functional and polished. Its dashboard read better even on low reasoning.
• Neither output was production ready. Both needed touch ups, wiring fixes, and styling cleanup. Plan for that no matter who wins.
• Cost was not the point. There are a hundred channels running the token spreadsheets. This was about capability per prompt.

The real lesson is that the model should match the artifact. Reach for Flash when you want a fast marketing surface that connects the pieces. Reach for GPT when you want clean functional polish on an internal tool. Stop picking one model for the whole stack.

How to Run This Test Yourself

If you want to repeat this without a PRD:

• Write a setup prompt with the business, the audience, the brand direction, and the artifacts. Nothing more.
• Ask the model to propose its own directory structure and brand decisions before writing code.
• Specify the port so you can pre-open preview tabs.
• Build one artifact at a time inside the same project.
• Click through the end to end flow. Submit forms. Type 9999 9999 into the phone field. See what breaks.
• Save the spec for the production build. Prototype on vibes.

I went in expecting GPT 5.5 to sweep this. It did not. Flash kept up where it mattered most for a small business marketing site, and that changes how I think about which model to reach for on which job.

Watch the full side by side build here: https://www.youtube.com/watch?v=G9gy1boCSB0