Kimi K3.1 Pro effort levels compared: low vs high vs max across 3 builds

Kimi K3.1 Pro has three effort levels: low, high, and max. The obvious question is whether the extra time and money on max buys you anything. So I ran the same three prompts on all three levels at the same time and compared the outputs side by side.

Short version: for business work, I'd go max every time, and the whole thing is cheap. For visual builds, Kimi didn't wow me. And every single effort level shipped text I could barely read.

Here's the full breakdown.

How I ran the test

I used Clearbench, an in-house tool I built for benchmarking models on every effort level side by side. It's still a work in progress, but it's been invaluable for these videos.

Every run is a clean start. No memory, no saved sessions, no skills, no plugins, no MCP servers, no hooks, no agent markdown files. Each level gets the same toolbox: read and write files, a shell with internet access, web search, and web fetch.

The important part is that all levels start at the same exact second. Each one gets its own sealed folder, and timing runs from the first file written to disk until the agent's last response. Then I open each build and see if it runs [00:38].

Build 1: the IKEA 3D assembly visualizer

The prompt: turn IKEA's page-by-page assembly instructions into an interactive, 3D, step-by-step visual lesson. I left every tech and creative stack decision up to the model.

Here's how the three levels landed:

• Low finished in 8 minutes 22 seconds and cost 59 cents. It works, but it skips steps and leaves out the screws.
• High finished in 18 minutes 37 seconds. Camera angle changes, an exploded view of all the parts, even the foot pads. Still no screws or tools.
• Max came in roughly 5 minutes after high. It finally shows the connecting screws, but the camera doesn't reposition the way I'd want, and the model doesn't look like the real product.

One thing I noticed across all three: they all went for the same left-and-right vertical menu layout. Nobody got creative outside that approach. I'm wondering if that's just Kimi's default front-end pattern.

Now, I have to be honest about my bias here. I use Opus 5.5 as my daily driver, and I ran this exact prompt on it. Opus one-shot the whole thing: animated assembly, a manual view that matches the IKEA pamphlet, a realistic view, every tool and screw, and how to turn each one [06:48]. So I'm spoiled. Kimi might do better with a more detailed prompt, but on the same prompt, it's not close.

Build 2: a full product launch funnel

This one is the practical test, and it's a prompt any business could borrow.

Some context. We're building Puddle, a plug-and-play agent harness that Clearmud will white-label for clients. The target customer is a company doing $5M to $20M in annual revenue. So I asked Kimi to build the whole launch package in one shot:

• A landing page and a lead funnel with a working backend
• A 90-day go-to-market strategy
• Research in the Meta Ads Library on businesses selling AI managed services, reverse-engineering the hook, the offer, the angle, the proof, and the call to action
• Master prompts based on those source ads, then its own
• Motion graphic ads

I also told it to check clearmud.ai for our brand persona and positioning [08:22].

### The numbers

| Effort | Time | Tokens | Cost | |--------|------|--------|------| | Low | 14m 26s | 2.35M | $1.00 | | High | a little under 2x low | almost 3x low | $2.29 | | Max | +10 min after high | 9.17M | about $3.28 |

### What each level built

Low was a solid foundation. A slick landing page with an early-access form, a 90-day strategy, ad research with source links back to Lindy and Relevance AI, four copy-paste master prompts, and basic motion graphics. It even built a downloadable lead-magnet PDF that I never specifically asked for. The motion graphics are meh and I'd change the styling, but for a one-shot at a dollar, that's impressive.

High nailed the above-the-fold. I get the hook, I get context, then I get an animated mockup. The ad research, the playbook, and the go-to-market plan each live on their own page, with a sticky layout I liked. There's a top menu banner, a little flowchart, and a detail where highlighted text turns green on hover that made me want to add green to our brand. Downside: it skipped the motion graphics entirely.

Max was the only run that gave me something that felt like a real backend. A "launch studio" with an overview, the ad research, motion creatives you can play side by side, the 90-day plan, and a calendar booking feature to request a call that neither of the other levels built [23:01]. The landing page itself was a bit plainer than high's, so I'd blend the two.

### What every level missed

Two things, and they were consistent.

First, nobody built the lead-capture backend. I asked for a working backend where I can see the leads, and none of the three delivered it. I kept wondering if it was my prompt, but this same prompt on other models doesn't miss that beat.

Second, the text. At 100% browser magnification, low, high, and max all rendered body copy that was painfully small [24:32]. I refused to zoom because I'm testing defaults. It's not a deal breaker, but I was shocked every level did it.

### My pick for this use case

Max, 10 times out of 10. When you're going through waves of ideation on a landing page, you don't care about a 15 or even 23 minute difference. The cost gap is a couple of dollars. Max gave me the most, so max wins.

I'll keep using Kimi for landing page ideation because it's good to see a fresh perspective on a website build.

Build 3: a 3D Sonic the Hedgehog clone

Last one was for me. I asked for a 1:1 recreation of Sonic in 3D, with the physics, momentum, and play style intact, plus a toggle between a first-person view and a third-person over-the-shoulder view. Personal use only.

• Low: 8 minutes 43 seconds. Honestly more impressive than I expected for the cost, and I've seen worse from more expensive runs.
• High: 26 minutes 58 seconds. The spin physics are off. Sonic rotates around the wrong anchor point instead of spinning in place.
• Max: just over 30 minutes. I hit right on my keyboard and Sonic goes left. Inverted controls, even in first person [30:47].

All three dropped the original side view entirely. I wanted the classic view with a hotkey to toggle into the new ones, and they all read that as "replace it." Could be my prompt, could be Kimi's interpretation.

Why I'm moving away from game clones as a benchmark

Just because a model is good at cloning a video game does not mean that model is good for everything. It might be the best model for visual-first or 3D work, and that still tells you nothing about how it handles a launch funnel or an agent task.

A lot of creators rank models on game builds because visual means clickbait. I'd rather focus on practical and creative use cases. I only ran Sonic because that game is part of my history, not because it makes a good thumbnail.

Final verdict on Kimi K3.1 Pro

• Visual builds: underwhelming. The IKEA visualizer missed the beat, though I admit Opus 5.5 set a very high bar.
• Launch funnel: very impressed with the UX and UI approach. That's where this model shines.
• Effort level: max, nine times out of ten, from an agent perspective. The time and cost difference is small enough that it doesn't matter.
• Still untested: Kimi K3.1 Pro as an agent brain. That's a future video.

Overall, it's a fast, affordable, efficient model, and I'd recommend testing it on your own use case before you write it off or go all in.

If you think I could have prompted Kimi more effectively, tell me how. I can't improve without your feedback, and it shapes what I build next.

Watch the full test here: https://www.youtube.com/watch?v=oOAnm-oy7e8