Claude Fable 5 vs. GPT-5.6 Sol vs. Kimi K3: Two Hands-On Benchmarks
Published on
Over the last few weeks I ran two mini-benchmarks putting the three best coding AI models in the world right now head to head: Claude Fable 5 (Anthropic), GPT-5.6 Sol (OpenAI), and Kimi K3 (Moonshot AI).
To be fair, both tests followed the same rules:
- Exactly the same prompt for all three models.
- All at the highest reasoning level possible: Max (Fable), Ultra (Sol), and Max (K3).
- All running on their official harness: Claude Code, Codex, and Kimi Code.
The first test was more creative, visual, and “from scratch,” while the second was an engineering job on existing code (redesigning my entire personal site).
Benchmark 1: “Space Discoverer,” a 3D space exploration game
The idea here was to create Space Discoverer, a 3D space exploration game running in the browser, focused on getting the best visual quality possible. I wrote a relatively generic prompt and let the models off the leash, basically saying “do whatever you think is best”.
Interestingly, all three converged on their own to the same stack: Three.js running on WebGL 2.
Below is my take on how each one did, plus execution time and token consumption.
Claude Fable 5 (Max)
It took 1h28, generating 257.7k output tokens. By far the one that generated the most code, more than triple the others.
Of the three, it delivered the best gameplay (with a more arcade feel that’s easy to control) and the prettiest visuals for the planets. The space soundtrack fit perfectly. The “warp speed” travel time isn’t fixed, which gives a real sense of distance.
One thing that caught me off guard (and that the others didn’t do) is that it created the black hole Gargantua, the one from the movie Interstellar, which caps your speed and makes the ship shake if you get too close. Nice touch!
The downside: the galaxies and nebulae work like “walls” and you can’t fly inside them, and the visuals stutter a bit.
GPT-5.6 Sol (Ultra)
It was the fastest, taking 48m27s, consuming 13.9M input tokens and 78.5k output tokens.
On gameplay, it chose a more realistic simulation style, which made the controls harder. Unfortunately it didn’t include music (which hurts immersion quite a bit), the planets came out kind of “poor,” and the background is a static, kind of fake nebula, with no stars like the others.
The galaxies, though, came out gorgeous, and it lets you fly inside them and inside the nebulae, unlike Fable.
Kimi K3 (Max)
It consumed 7.5M input tokens and 84.3k output tokens in ~1 hour.
Just like GPT, the planets didn’t come out very pretty. But it added a soundtrack like Fable did, and the sound effects of the ship accelerating came out excellent. It synthesized every sound in the game purely using the WebAudio API.
Where it really impressed me was the look of the nebulae: flying the ship through them feels like flying inside a real cloud of stardust, full of newborn stars. The Sombrero galaxy also came out beautiful, looking like a quasar.
See the three games in detail
You can also play the three games in your browser and see the exact prompt I used.
Benchmark 2: a complete redesign of this site
In the second test, I completely changed the type of task: instead of creating something from scratch, I asked the three models to work on a real codebase, my own personal site, where you’re reading this article right now.
The request was basically for them to analyze the entire site (layout, visual hierarchy, typography, colors, navigation, accessibility, performance, etc.) and, based on that, implement a complete redesign directly in the code, with my identity, and not a generic portfolio template.
The execution times were already a first interesting data point: Fable took 35 minutes, K3 took 50 minutes, and Sol took 1h10.
The most curious part is that all three independently ended up in practically the same direction: a “warm paper + ink blue” background, editorial serif typography, and an element highlighting my 14 languages.
Two of them (Fable and Kimi) named this concept with exactly the same name: “Editorial Polyglot” (“Editorial Poliglota” in Portuguese). Could that be a sign of distillation?
They all also removed heavy dependencies (the Tailwind CDN and the entire Lucide library), swapping them for much lighter local CSS, and followed the same accessibility guidelines.
Sol went down a slightly different path, with the concept “global AI field notes” (AI becomes IA, Barcelona becomes Brazil), and focused A LOT on performance optimization: it cut the deploy size in a monstrous way, from about 382MB to 135MB, heavily changing the format and size of the images and videos.
My honest opinion on each one
Claude Fable 5: I really liked the color palette and the visual style. It came out looking sophisticated and, even though all three picked a similar palette, I thought it was the one that applied it best in practice.
GPT-5.6 Sol: I found it amazing that I only mention my project Sotaque in a random blog post (and on GitHub), and it went and understood what the project was, found the logo, and gave it a highlight on the site. A level of reasoning that impressed me. I also liked how it structured the highlights. However, I found the texts it wrote EXTREMELY cheesy, with a “coach” vibe far from my style, plus several little block and text positioning slip-ups. It would need a few more rounds of back and forth.
Kimi K3: it came out similar in style to Fable, but a bit blander and less polished in the overall finish (the posts and articles section isn’t centered like the rest, for example).
Compare the three versions of the site
In the video below I compare the site as it is today with the three generated versions:
The exact prompt I used for all three is here.
Wrapping up
Putting the two tests together, the pattern is clear: nobody won at everything.
Fable 5 delivered the best visual “taste” in both tasks, but generated A LOT more code than the others (which can be a big problem if you’re paying for tokens through the API). Sol did really well at reasoning over an existing codebase and at performance optimization, but slipped on aesthetic finish and writing tone. Kimi K3 sits in the middle, with really high peaks in specific spots, like the synthesized audio and the game’s nebulae.
They all worked well, and that’s exactly why it’s worth running this kind of test on your own tasks, and not just looking at benchmark tables.
So: which of these versions did you prefer, and why?