Which ChatGPT 6.1 Terra Effort Level Should a Small Business Use? I Tested All Six

If you're set on using ChatGPT 6.1 Terra to rebuild your business website, I'll save you the time: use Max.

I ran all six effort levels (low, medium, high, extra high, max, and ultra) through two practical small-business builds. Max was the only one that listened to the full instruction and rebuilt a page for every page on the live site. And it still only cost 24 cents.

Here's how I tested it, what each level got wrong, and the one thing I'd change before using Terra for this kind of work again.

How the test stays fair

We built an in-house bench for these videos, and I want to be clear about how it works, because a sloppy setup makes for a sloppy comparison.

• One prompt, sent to all six effort levels at the same exact time.
• A clean start: no memory, no saved sessions, no skills, plugins, MCP servers, or hooks.
• The same toolbox for every run: read and write files, a shell with internet access, web search, and web fetch.
• A sealed folder for each run so there's no cross-contamination.

That last one matters more than it sounds. In past tests, when models worked out of a shared directory, they'd go look at the root folder and reference code that was already there. Not a fair test.

Timing runs from the first file written to disk until the agent's last word. We convert that into an equivalent API spend, then check one thing: does it run?

Build 1: rebuild a website that looks like 2026, not 2010

The target was a local business I found randomly on Google Maps. No affiliation. The site isn't horrible, but it's clearly from the previous decade. What I liked about it was the brand: a warm, welcoming orange and plenty of creative assets to work with.

The prompt asked Terra to visit the link, pull the logo, brand persona, target audience, industry, and creative assets, then make its own creative and tech stack decisions.

### Terra is shockingly fast and cheap

I didn't need to fast-forward. Medium finished first. Ultra was already showing a preview. Low, oddly, was the slow one. Max took the longest.

All six finished in roughly one minute. Five of the six came in under 10 cents each. Max, the longest run, cost 24 cents.

For context, Terra is not the best model I've tested at common builds, game dev builds, or 3D builds. Where it's strongest in my tests is computer use, video editing, and powering an agent harness. But that price point is exactly why I wanted to try business builds anyway. A small or mid-size company on a budget would find this very appealing, right?

### What the first run got wrong

Walking down the list:

• Low pulled the logo but not the color palette. It defaulted to the same generic-looking modules GPT likes, and it was a basic one-page site. It would need a lot of follow-up prompts.
• Medium got closer, but the brand came out red instead of orange. Still missing the beat on pulling the palette from the source site.
• High defaulted to the same look and the same wrong color.
• Extra High was the first level that looked like the brand. Orange throughout, with one green hover state I wasn't thrilled about.
• Max added pop-up modules and more brand assets. One problem: the "get a free estimate" button linked out to the current live page.
• Ultra looked a lot like the others and also linked out to the old site.

Every single one of them built a one-page site. The live site has separate pages for each service. That's partly on me, which brings me to the rerun.

The one sentence that changed the test

My first prompt never told Terra to recreate every page. So I added one sentence to the end:

"I want you to analyze every single subdirectory, every single page on the current live website and rebuild this one for one, re-imagined with this newer look."

Then I ran the bench again.

This time, nothing was done at 30 seconds. The models now had to pull the sitemap and crawl every page. Medium and High again finished before Low, which is still a weird discovery to me.

### Only Max followed the instruction

Low looked a lot better on the rerun and pulled the orange, but it still didn't build out the pages. Clicking a service opened my email app. Medium, High, Extra High, and Ultra all did the same thing: one top-down landing page, no separate service pages.

I started wondering whether this was a limitation of our bench rather than the model, so I checked. It's the model.

Max was the only effort level that listened [15:33]. It rebuilt every subdirectory one for one, with little orange details everywhere. It became my favorite output of the whole video.

So, one out of six. If you're set on Terra for a website rebuild, use Max.

Build 2: a full product-launch kit for Puddle

The second prompt was bigger. I asked for a landing page that doubles as a lead funnel, a working backend to capture leads, a go-to-market strategy, and ads modeled on ones already working in the industry.

The product was a fictional Clearmud offering called Puddle: a plug-and-play, white-label AI agent harness for busy professionals who don't want to learn a new framework. The prompt pointed Terra at clearmud.ai for brand persona, style, and positioning, and described Clearmud as a YouTube channel that works with brands and provides managed AI services to companies with $5M to $20M in annual revenue.

I told it to start with research: search the Meta Ads Library for similar businesses and AI-oriented ads selling managed services, reverse engineer why they work (the hook, offer, angle, proof, and call to action), build master prompts from those source ads, then create its own creative, motion graphics included. It could use whatever skills it found useful.

### Four minutes, 24 cents, and no backend

The longest run was 4 minutes and 6 seconds at 24 cents. Finish order was High, Medium, Low, Ultra, Extra High, then Max. Extra High took the second longest but cost less than Low and Ultra. All six went with Node. Max took the most turns and still landed at 24 cents.

Then I opened them up:

• Low pulled something close to our blue, made an odd logo adaptation, and gave a decent lead capture. None of the other artifacts I asked for.
• Medium pulled our real logo and had a similar structure. Not a bad foundation, but no links to the rest.
• High went straight to a green and cream default look that has nothing to do with our brand. It lost my trust immediately.
• Extra High didn't feel like our brand either, with a very basic UX approach.
• Max had the best above-the-fold, a nice mockup, a strong call to action insert, and a good fact section. Still no links to the backend.
• Ultra missed our color palette, so I stopped there.

Not one effort level delivered a backend I could click into. Other models I've tested handle that request fine, so I think this was simply too big an ask for Terra as prompted.

Terra needs a PRD

Here's the pattern across both builds. Terra cuts corners that other models wouldn't cut. It does a decent job, don't get me wrong, but it's so quick that it skips things.

With a lot of my recent benchmarks, especially with Opus and more intelligent models, I've found that a PRD (product requirements document) isn't needed anymore. Terra might be the exception. With a PRD, you're not testing the model's creativity or its tech stack choices. You're handing it every specific page, every tab, how you want it to look, and what to build on. For these kinds of builds, I have a feeling the results would be far better. That's giving me an idea for another video.

Key takeaways

• If you're choosing Terra for its cost, speed, and efficiency, use Max. 10 times out of 10. It's cheap enough, fast enough, and it gave the best outputs in both builds.
• Give it all the context up front. One added sentence about rebuilding every subdirectory changed what Max delivered.
• Prep a PRD before a business build with Terra. It benefits from one more than the smarter models do.
• Don't expect a full backend from a single open-ended prompt. Break that out.

If you disagree with me on Max, drop a comment and tell me which effort level you thought was best. And if you want to see me test models from inside different agent harnesses, we have an agent bench web app coming very soon. Let me know you're interested.

Watch the full test here: https://www.youtube.com/watch?v=LcqZyWqdVrA