Blog Article

Which Gemini 4 Carbon Effort Level Should You Use? I Tested All Three

I ran three one-shot builds through Gemini 4 Carbon on low, medium, and high effort: a paintball game, a 30-second promo video, and a landing page. High effort won

I ran the same three prompts through Gemini 4 Carbon on low, medium, and high effort. One first-person paintball game, one 30-second motion graphics promo, and one landing page that doubles as a lead funnel.

Short answer: high effort gave me the best result in all three tests. It wasn't much slower, and it stayed cheap. In one test, high cost less than medium.

Let me show you how I got there.

How I kept the test fair

I ran everything through an in-house benchmarking tool I made for these videos. The setup matters, so here's how it works:

  • One prompt, clean start. No memory, no saved sessions, no skills, no plugins, no MCP servers, no hooks, no CLAUDE.md, AGENTS.md, or any markdown files.
  • Same toolbox. Every run can read and write files, use a shell with internet, search the web, and fetch web pages.
  • Same start time. All three effort levels kick off together.
  • Sealed folders. Each run builds in its own folder, so nothing leaks between tests.
  • Cost from the source. Tokens come from the CLI's own report, cross-referenced with retail API prices.

These are one-shot prompts. No follow-ups, no fixes. What you see is what the model built on the first try.

Test 1: A first-person paintball game

I'm a big paintball guy, so this one was personal.

On the scorecard, low finished fastest and was the cheapest. Medium and high were close on time. The cost went up about $1 with each effort level, which is not bad at all.

How they played:

  • Low: No invert-mouse option, and when I got hit I had to walk out myself. Crouch worked. Not bad for low effort, but nothing too impressive.
  • Medium: It had an invert option in the settings and felt a lot smoother. My arms detached at one point, and the bunkers weren't that realistic, but it was playable.
  • High: I couldn't find an invert option, which was a bummer. But the paint splash and the bounce on the bunkers looked nice.

Now, let me be clear. Would I choose Carbon over Opus and Fable for this? Not in its current form.

But the high-effort build only cost $3.40 [05:41]. For a playable game from one prompt, that's kind of amazing. If I took a multi-layered approach and built out proper 3D assets for the guns and bunkers, I think we'd get a much better result. That just wasn't the point of this test.

Test 2: A 30-second promo video

I didn't have a master prompt ready for this one, so I vibed it. I asked Carbon to build a promotional video for Claude's Fable 5.2, go fetch the creative it needed from Anthropic, do a better job than they did, include sound, and have some fun. Then I set the length to 30 seconds.

Low cost $1.77. The animation was not bad at all and it went above and beyond what I expected. The problems: the voice-over was bad enough that I removed it, and the video had a weird stutter. Still, for $1.77 and something we can fine-tune, that's not horrible. I'm not a creative professional. If I had given it a proper shot list, I have a feeling low would keep surprising us.

Medium was a lot smoother, with no stutter. I would have loved to see it pull in the Claude mascot, but I didn't ask for that. The more time you spend on the upfront prompt, the better the result.

High is where it got interesting. It built the video, which is the only thing I asked for, and then it built a landing page for it too [08:51]. It also created its own chart based on the benchmark chart screenshots on Anthropic's page. Color me impressed. I would choose high all day.

Then I looked at the scorecard. The runs finished in the order you'd expect: low, medium, high. But high was cheaper than medium. Are you kidding me?

After this test, Gemini 4 Carbon earned a spot in my ideation workflow. I'll run the same prompt across Fable, Opus, and now Carbon. For visual and motion graphics work, I like it more than the GPT models.

Test 3: A landing page with no PRD

This approach works for any small to mid-size company or entrepreneur who already has a website live and needs a landing page for a product.

I pointed Carbon at clearmud.ai and told it to get a feel for the brand positioning, color palette, typography, and assets. Then I asked for a landing page that also serves as a lead funnel for a new product: a white-label-friendly AI agent harness that small to mid-size companies can plug into their environment.

Normally I'd write a PRD, or product requirements document, for this. I skipped it on purpose. I wanted to see if Carbon could go out and think for itself. In my recent tests, 6.1 Terra did a horrible job at that. It needs a PRD to build things.

The scorecard:

| Effort | Time | Cost | |--------|------|------| | Low | 17 minutes | $1.26 | | Medium | 19 minutes 51 seconds | $1.29 | | High | 25 minutes 32 seconds | Not stated in the video |

Three cents between low and medium. I was shocked.

The pages themselves:

  • Low: It pulled some effects from our site, the table looked slick, and it added a "customize your harness manifest" feature. But everything above the fold was a horrible one shot, and the nav header made me want to barf. I wouldn't choose low for a landing page.
  • Medium: Better, with no gibberish text and a slick table. There was way too much crammed up top, though. It's a pretty solid foundation if you want to build from it, but it needs fine-tuning.
  • High: The best nav header of the three, less condensed, and a great looking chart. It did a better job listening to instructions.

Would I launch the high-effort page as is? In all reality, no. It needs some TLC. But it's the clear winner.

So which effort level should you use?

High. Across the board.

  • Paintball game: best splash and bounce effects for $3.40
  • Promo video: the video plus a bonus landing page and a custom chart, for less than medium cost
  • Landing page: best header, best layout, about eight minutes slower than low

High is so dang cheap, it's not that much slower, and it gave me the best results in all three tests. If I'm building web apps or landing pages with Carbon, high is what I'd pick.

Low still has a place. If you want a quick, cheap first pass on motion graphics that you plan to fine-tune, $1.77 gets you a lot.

And remember, these were loose one-shot prompts. A real PRD or a shot list would push every one of these results further. All you got to do is try.

What should I test next?

If you want to see Gemini 4 Carbon go up against Argon, Fable, or Opus, drop a comment on the video and tell me specifically what you want me to build. I'll add it to the queue.

Watch the full test here: https://www.youtube.com/watch?v=ucXi7Cq0Lrw

Watch the full walkthrough on YouTube.

Watch on YouTube