Kimi K3 vs the Frontier
Ten prompts. Six models. Every build published live and unedited.
Kimi K3▲
GPT-5.6 Sol◉
Opus 4.8◆Kimi K2.7△
GLM 5.2⬢Fable 5✳Kimi K3 just came out. It's Moonshot's new open model, and the pitch is frontier-level performance for a fraction of the price. So I put it to the test.
I ran it against GPT-5.6 Sol, Opus 4.8, Fable 5, GLM 5.2, and the last Kimi, K2.7, all on the same prompts. Some are real builds, like a 3D gadget, a weather app, a fighting game, and a luxury watch page. The rest is knowledge work, like a Tarantino trailer, a motion graphic, and a landing page rewrite.
No retries, nothing cleaned up, and every build is live. The chapters are on the left. Copy any prompt, run it yourself, and see the builds side by side.
Build an interactive 3D command-center gadget
What to watch: Orbit the device, then actually operate it. Does every knob, switch, and slider drive something on the screen, or are half of them decorative?
Kimi K3▲
GPT-5.6 Sol◉
Opus 4.8◆Kimi K2.7△
GLM 5.2⬢Build a weather app with its own forecast model
What to watch: Search a few cities. Is it pulling real live data and running its own forecast next to the official one, or just skinning a current-conditions widget?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Build an MP3-to-MIDI converter
What to watch: Play the source, then play the MIDI. Does the transcription actually sound like the melody, or is it random notes wearing a piano roll?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Build a 2D fighting game
What to watch: Play a full round. Do the combos register, does blocking matter, and does the state machine hold, or can you attack straight through a hitstun?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Build a luxury watch product page (3D)
What to watch: Spin the watch up close. Do the metal, crystal, and dial read like a real studio product shot, or like flat 3D primitives?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Fable 5✳Build a Counter-Strike 2 browser clone
What to watch: Does it play? Check movement, hit detection, and whether the map has real collision instead of a skybox you fall straight through.
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Fable 5✳Write an Isaacson-style biography intro
What to watch: Read the sources note. Is every biographical claim grounded in something real, or did it invent a hometown and a backstory?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Write + shot-list a 30-second Tarantino trailer
What to watch: Watch it through. Does the dialogue land like peak-90s Tarantino, and does the shot list read as one coherent scene or seven unrelated clips?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Author a 20-second motion-graphics explainer
What to watch: Watch the twenty seconds. Do the beats develop with intentional motion, or is it a static slide with a fade?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢Rewrite a real workshop landing page
What to watch: Read the copy out loud. Does it dig into a founder's real pain and sound human, or is it generic AI slop with 'community' stapled on?
Kimi K3▲GPT-5.6 Sol◉Opus 4.8◆Kimi K2.7△GLM 5.2⬢So is K3 beating Opus and Sol across the board, the way the benchmark charts want you to believe? Honestly, probably not. On the work that pays the bills, the copywriting and the harder builds, Opus and Sol still edge it out. But look where K3 landed on the creative and the interactive: first on the gadget, first on the motion graphic, right in the mix on the fighting game and the hallucination test. Next to K2.7, this is a real step up.