I built the same five interactive apps twice: once with Moonshot AI's Kimi K3, once with Claude Fable 5. I hid the model names until the very end. No labels, no hints, just two outputs side by side for each app.
In this article, I walk you through every round exactly as I judged it, then break down what Kimi K3 actually is, what it costs, and how its official numbers stack up against Fable 5.
Key Takeaways
How I Set Up the Blind Test
I gave both models the same five prompts and asked for rich, interactive UI apps, the kind of build where a model can't hide behind a static screenshot. I ran Fable 5 through Claude Code at medium reasoning effort. I ran Kimi K3 through an open-source coding CLI at max effort.
I never labeled which output came from which model while I recorded my reactions. I revealed the answer only after judging each round on its own. If you want the exact prompts I used, grab them from the link in the description box, or pull the full prompt library from promptslove.com.
Round 1: 3D Floor Plan Walkthrough
See the Pen 3D Floor Plan - Fable by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
See the Pen 3D Floor Plan - KK3 by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
I fed both models a floor plan image and asked for an interactive 3D walkthrough. Both outputs gave me three views: the full house, a top-down plan, and a first-person walkthrough. Both nailed the dollhouse view, the room-by-room navigation, and the hover-to-inspect details showing wall dimensions, wood type, and furniture.
The difference showed up in the walkthrough view. One model's walkthrough camera stayed stuck outside the walls. I could not step into a room. The other model let me walk straight into any room, mouse over furniture, and see live dimensions and accessory labels.
"Absolutely stunned by the output because, without errors, it was able to do such an interactive app."
The model with the working walkthrough was Kimi K3.
Build BluePrint3D — an interactive 3D floor-plan previewer. A furnished apartment/house
shown in 3D; hovering a room highlights it and shows a detail tooltip/panel.
DELIVER self-contained blueprint3d.html. Three.js r128 + GSAP + Google Fonts via CDN allowed.
RUN AT: reasoning_effort max.
═══════════════ THE 3D FLOOR PLAN ═══════════════
Model a complete single-level home in 3D from Three.js primitives (no external models):
- A realistic multi-room layout (e.g. 2-bed apartment): Living Room, Kitchen, Master
Bedroom, Second Bedroom, Bathroom, Hallway, Balcony. Define via a data structure
(rooms with floor polygons, wall segments, doors, windows).
- FLOORS: each room's floor as a slab with a material hinting its type (wood for living,
tile for kitchen/bath, carpet for bedrooms) — different colors/textures per room
- WALLS: extruded from wall segments, with door GAPS and window openings cut in
- DOORS + WINDOWS: simple framed geometry in the openings
- FURNITURE (low-poly, built from primitives): sofa + coffee table + TV unit (living);
counters + island + fridge + stove (kitchen); bed + wardrobe + nightstands (bedrooms);
sink + toilet + shower (bath). Enough to make each room read clearly.
- Soft lighting: ambient + a key light; soft shadows; a pleasant, clean architectural look
(think a polished real-estate 3D tour, semi-stylized, not photoreal)
═══════════════ CAMERA / VIEW ═══════════════
- Default: a nice angled "dollhouse" isometric-ish view of the whole plan
- Orbit + zoom + pan (drag/scroll/right-drag)
- View presets: Top-down (true floor-plan view), Dollhouse (angled), and a per-room
"walkthrough" framing when a room is selected
- Toggle: walls full-height vs cutaway/short (so top-down reads like a plan)
═══════════════ THE MOUSEOVER DETAIL (the core requirement) ═══════════════
- Raycast on mouse move: the room under the cursor HIGHLIGHTS (floor/walls tint or glow,
others dim slightly) with a smooth transition
- A TOOLTIP follows the cursor (or a fixed side panel updates) showing that room's detail:
room name, area (m² / ft²), dimensions, floor type, number of windows, a short
description, and maybe an icon set (bed/bath/etc.)
- On CLICK a room: camera smoothly frames it, a fuller INFO PANEL slides in with all
details + a small furniture list, and the room stays selected/highlighted
- Hovering furniture (optional nice-to-have): show the item name in the tooltip
- Smooth enter/leave transitions on all highlights + tooltips
═══════════════ THE RICH UI ═══════════════
- Left dock: room list (click to select/fly-to), each with area + icon
- Top bar: view preset buttons (Top / Dollhouse / Walkthrough), wall-height toggle,
units toggle (m²/ft²), a measure/dimensions toggle
- Right panel: selected-room details (area, dimensions, floor, windows, furniture list,
description) with animated entrance
- A total summary: total area, room count, bed/bath count (like a listing header)
- Clean architectural aesthetic: blueprint-blue + white + a warm accent, crisp type,
subtle grid background on the UI
- A legend for floor-type colors
DATA STRUCTURE: define rooms as data (name, floor polygon points, height, floor_type,
windows, area, description, furniture[]) and BUILD the 3D from that data — so hover/
click can map a mesh back to its room record for the detail panel.
VERIFY: full home renders in 3D with distinct rooms + furniture; orbit/zoom/pan work;
hovering a room highlights it + shows a detail tooltip; clicking frames it + opens the
info panel; view presets + wall-height + units toggles work; room list navigates;
areas/dimensions are shown; 60fps.
Output the complete file, no truncation. The room hover-highlight + live detail readout
is the central feature — make it crisp and satisfying.Round 1 winner: Kimi K3.
Round 2: Browser Racing Game
See the Pen Neon Circuit Game - Fable by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
See the Pen NEON Circuit Game - KK3 by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
Next, I asked for a browser-based racing game with difficulty settings, a leaderboard, power-ups, and a vehicle picker. Both apps let me pick "hard" difficulty, choose a vehicle called the Interceptor, and race through a track with obstacles and nitro boosts.
One build ("Neon Circuit") had noticeably sharper graphics, working sound effects on nitro pickups, and obstacles that visibly cost me health and speed on contact. The other build started faster with less setup friction, but the graphics and sound were a step down.
I preferred the left build for its graphics and audio feedback. That build turned out to be Kimi K3 again, built entirely in plain HTML, CSS, and JavaScript.
Build Neon Circuit — a polished arcade racing game with a AAA-quality animated UI layer.
DELIVER a single self-contained neon_circuit.html. All HTML/CSS/JS embedded. Opens by
double-clicking. Google Fonts allowed; no other external resources.
RUN AT: reasoning_effort max.
═══════════════ THE GAME ═══════════════
Top-down / pseudo-3D endless neon racer on a dark synthwave grid.
Canvas 960×600, 60fps, requestAnimationFrame.
GAMEPLAY:
- Player car on a multi-lane highway (3-5 lanes), neon wireframe aesthetic
- Steer left/right (A/D or arrows) between lanes; accelerate/brake (W/S)
- Dodge traffic cars + obstacles; collect energy orbs + boost pads
- Speed increases over time; score = distance + orbs
- A pseudo-3D perspective road (scanline scaling of lane markers toward a horizon)
- Nitro boost meter (fills from orbs, spend for a speed surge with motion blur)
- 3 lives / shield system; crash = lose shield, screen shake, particle burst
- Difficulty ramps: more traffic, faster scroll, hazard types unlock
VISUAL / JUICE (this is where K3 must excel):
- Synthwave palette: deep purple/blue bg, hot-pink + cyan neon, sun-grid horizon
- Animated scrolling grid floor with perspective
- Glowing neon car sprites (drawn procedurally with layered glow)
- Particle systems: exhaust, orb pickup sparkle, crash debris, boost trails
- Screen shake, chromatic-aberration flash on crash, speed-line streaks at high velocity
- Parallax star/city background layers
═══════════════ THE RICH UI LAYER (the K3 frontend test) ═══════════════
This is the differentiator — a genuinely premium, fully animated game UI:
MAIN MENU:
- Animated title with neon flicker + glow
- Menu items with hover states (slide, glow, sound-ready hooks)
- Animated background (the game world idling behind a blur/vignette)
- Panels: Play, Garage, Leaderboard, Settings, How to Play
GARAGE (car select):
- 4-6 selectable cars, each with stats bars (speed/handling/boost) that animate
- 3D-ish car preview that rotates (CSS 3D transform or canvas)
- Unlock states, selection animation
IN-GAME HUD:
- Animated speedometer (arc gauge), nitro meter, score counter (odometer roll),
lives/shield icons, mini lap/distance tracker, combo multiplier popups
- Damage/boost vignette effects layered over the canvas
PAUSE / GAME OVER:
- Frosted-glass overlay panels, animated stat reveal (count-up),
"new high score" celebration with confetti/particles
- Smooth transitions between all screens (no hard cuts)
SETTINGS:
- Volume sliders, difficulty, control scheme, graphics quality toggle,
color-theme picker (swap the neon palette live)
LEADERBOARD:
- Top 10 (localStorage), animated row entrance, current-run highlight
ALL UI must have: smooth CSS transitions, hover/focus states, entrance animations,
and a cohesive design system (shared color vars, consistent spacing, one type scale).
It should feel like a shipped indie game's menu, not a prototype.
STATE MACHINE: MENU, GARAGE, LEADERBOARD, SETTINGS, HOWTO, PLAYING, PAUSED, GAMEOVER
with animated transitions between all states.
PERSISTENCE: localStorage for high scores, selected car, settings, unlocks.
CODE ORGANIZATION (single file, but structured):
Clear class separation: Game, Player, Traffic, Orb, ParticleSystem, Road (perspective),
UI (screen manager), HUD, Audio (Web Audio synth SFX — engine hum, pickup, crash, boost),
Storage. Comment the sections.
AUDIO: synthesize all SFX via Web Audio API (no audio files) — engine drone that pitches
with speed, pickup chime, crash noise, boost whoosh, menu clicks.
VERIFY: menu→garage→play flow works; steering + collision + orbs + nitro all work;
crash reduces shield + shakes; game over shows animated stats + saves high score;
all UI screens reachable and animated. 60fps target.
Output the complete single HTML file, no truncation. The UI layer must be as polished
as the gameplay — that's the point of this build.Round 2 winner: Kimi K3.
Round 3: Live 3D Solar System
See the Pen Live Solar System - Fable by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
See the Pen Orrey - Live Solar Simulator - KK3 by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
For round three, I asked for a live, real-time 3D solar system with a "grand tour" mode that cycles through the planets automatically. Both apps rendered a 3D scene with mouse-controlled zoom and pan, a planet picker, and a details panel that updates as you select each planet.
One build's grand tour camera glitched when zooming into a planet, then snapped to a different, less polished planet view. The UI on that build also felt less refined overall. The other build's grand tour ran cleanly through Mercury, Venus, Earth, and beyond, with a UI I found more intuitive.
Build Orrery — a live, interactive 3D solar system with a rich information UI.
DELIVER self-contained orrery.html. Three.js r128 + GSAP + Google Fonts via CDN allowed.
RUN AT: reasoning_effort max.
═══════════════ THE 3D SOLAR SYSTEM ═══════════════
A real-time WebGL solar system the user can explore.
SCENE:
- The Sun at center: emissive sphere with a glow/corona (shader or layered sprites),
subtle surface noise animation, a point light illuminating the planets
- All 8 planets (Mercury→Neptune) as spheres with procedurally-generated or
canvas-painted textures (banded gas giants, rocky terrestrials, blue Earth with
cloud layer, ringed Saturn with a real ring geometry, tilted Uranus)
- Each planet orbits the Sun on its real relative orbital period (scaled to be watchable)
and rotates on its axis (with correct-ish axial tilts)
- Moons for Earth (Luna) and Jupiter (the 4 Galileans) at minimum
- Saturn's rings (ring geometry with a gradient/alpha texture) and a faint asteroid belt
(instanced particles between Mars and Jupiter)
- A starfield skybox (thousands of points or a star texture sphere)
- Orbit path lines (faint ellipses) toggleable
SCALE MODES (a real feature):
- "Realistic" (true-ish relative sizes/distances — planets tiny, far apart) vs
"Compact/Educational" (exaggerated sizes + compressed distances for a nice view).
- Toggle animates smoothly between the two.
CAMERA / INTERACTION:
- Orbit + zoom (drag/scroll). Smooth momentum.
- Click a planet → camera smoothly flies to it (GSAP), focuses/frames it, and the
info panel populates with that planet's data
- "Follow" mode: lock the camera to a planet as it orbits
- Free-fly toggle for exploration
- Time controls: play/pause, speed slider (1×…100000×), "reverse", "now" reset
═══════════════ THE RICH UI ═══════════════
A polished, sci-data overlay (think NASA-eyes / stellarium aesthetic):
- Top bar: title, current simulation date/time (advances with the sim), time-speed control
- Left dock: planet selector list (click to fly-to), layer toggles (orbits, labels,
asteroid belt, moons, realistic/compact scale)
- Planet INFO PANEL (on select): name, type, diameter, mass, gravity, day length,
year length, distance from Sun, moons count, temperature, a fun fact, and a small
stat-bar comparison vs Earth. Animated entrance. (Hard-code accurate planetary data.)
- Bottom: a timeline / orbital-position mini-map (top-down radar of the system showing
where each planet currently is; click to select)
- Labels floating on planets (billboarded, toggleable, fade by distance)
- A "tour" button: an auto cinematic that flies through the planets in order with the
info panel narrating each
VISUAL QUALITY:
- Bloom on the Sun + bright highlights; soft ambient + the Sun's point light
- Nice materials (roughness/metalness), atmospheric rim glow on Earth/gas giants
- Smooth 60fps; cap pixel ratio; dispose properly; reduced-motion fallback
DATA: hard-code accurate data for all 8 planets (diameter, mass, gravity, orbital period,
rotation period, distance, moons, temp, axial tilt) so the info panels are real.
STRUCTURE: scene/lighting setup, celestial-body factory (buildPlanet with texture gen),
orbital mechanics update, camera controller (orbit + fly-to + follow), UI/info panel,
radar minimap, time controller, animation loop.
VERIFY: planets orbit + rotate at correct relative rates; click-to-fly-to works; info
panel shows accurate data; time controls speed/pause/reverse; realistic↔compact toggle
animates; orbit lines + labels + asteroid belt toggle; Saturn has rings; starfield;
Sun glows; 60fps.
Output the complete file, no truncation. This should look like a real interactive
astronomy tool with a beautiful UI — a portfolio-grade showcase.I liked the cleaner build here, which was Fable 5 running at medium reasoning effort.
Round 3 winner: Fable 5. Score after three rounds: Kimi K3 2, Fable 5 1.
Round 4: Global Maritime Trade Map
See the Pen Sealanes - Gobal Maritime Trade Map - Fable by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
See the Pen Sealanes - Trade Arteries - KK3 by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
Round four asked for an interactive world map showing major maritime trade routes, with hover details on cargo type, transit days, and distance for routes like the Strait of Hormuz and the trans-Atlantic corridor. Both apps delivered neon-lit shipping lanes on a world map, plus a rotating or flat globe animation.
Both builds handled the hover tooltips and cargo data well. The difference was instruction-following: I asked the globe to stop rotating on command. One build kept spinning no matter what. The other stopped exactly when told.
"The instruction following is not there in the Kimi K3. Anyway, I love the output from Kimi K3, but I like Fable Five better."
Build SeaLanes — an interactive visual world map showing all major global shipping routes.
DELIVER self-contained sealanes.html. Use D3.js (v7) + TopoJSON world atlas via CDN, plus
Google Fonts. (D3 for the geographic projection + world land shapes.)
RUN AT: reasoning_effort max.
═══════════════ THE MAP ═══════════════
- A world map using a D3 geographic projection (default: a nice equirectangular OR
an orthographic GLOBE you can drag to rotate — offer BOTH via a toggle: "Flat" vs "Globe")
- Load world land polygons from a TopoJSON world atlas (e.g. world-atlas countries)
- Ocean, land, and country borders styled cleanly (a premium cartographic look —
think a dark "trade map" aesthetic: deep navy ocean, muted landmasses, glowing routes)
- Graticule (lat/long grid) faint in the background
- Zoom + pan (d3.zoom); on the globe, drag to rotate + scroll to zoom
═══════════════ THE SHIPPING ROUTES (the data core) ═══════════════
Hard-code a comprehensive dataset of the world's MAJOR maritime shipping routes as
arrays of [lng,lat] waypoints, including at minimum:
- Trans-Pacific (East Asia ↔ US West Coast)
- Asia ↔ Europe via the Suez Canal (the big one: China→Singapore→Suez→Rotterdam)
- Asia ↔ Europe via Cape of Good Hope (the Suez alternative)
- Trans-Atlantic (US East Coast ↔ Europe)
- Europe ↔ South America
- The Panama Canal routes (Atlantic↔Pacific)
- Middle East oil routes (Persian Gulf → Asia; Gulf → Europe)
- Intra-Asia, Australia ↔ Asia, Africa coastal routes
- The Northern Sea Route (Arctic) as an emerging lane
Draw each route as a smooth GREAT-CIRCLE-style curved line following the projection
(use d3.geoInterpolate or path along the sphere so lines curve naturally on the globe).
Route styling: glowing gradient lines; animated "flow" dashes moving along each route to
show direction of travel; thickness by relative traffic volume.
CHOKEPOINTS: mark and label the critical maritime chokepoints with special markers +
pulsing highlights: Suez Canal, Panama Canal, Strait of Malacca, Strait of Hormuz,
Bab-el-Mandeb, Bosphorus, Gibraltar, Cape of Good Hope. Hovering a chokepoint shows a
detail card (why it matters, approx % of global trade).
PORTS: plot the top ~20 container ports (Shanghai, Singapore, Rotterdam, LA/Long Beach,
Ningbo, Shenzhen, Busan, Dubai, Antwerp, Hamburg, etc.) as markers sized by throughput.
═══════════════ INTERACTION & RICH UI ═══════════════
- Hover a route → it highlights (others dim), a tooltip/card shows: route name, endpoints,
approx distance, transit time, key cargo, chokepoints it passes
- Hover a port → name, country, annual TEU throughput, rank
- Hover a chokepoint → its detail card
- Click to pin a route/port and keep its details in a side panel
- LEFT UI dock:
* toggle groups: show/hide routes by category (container, oil/tanker, bulk),
ports, chokepoints, graticule, country borders
* a route list (click to highlight/zoom to a route)
* Flat ↔ Globe projection toggle
* animation speed for the flow dashes
- LEGEND: line color/thickness meaning, marker types, traffic-volume scale
- Search box: find a port or route by name → highlights + zooms
- A subtle stats header: total routes shown, total ports, "global trade" flavor stats
VISUAL QUALITY:
- Dark, premium "geopolitics/trade-intelligence" aesthetic (deep navy ocean, glowing
cyan/amber routes, clean sans type). Smooth animations. Everything should look like a
high-end data-journalism interactive (NYT/Financial-Times-grade).
- Animated flow along routes (marching dashes / moving particles) is essential — it makes
the map feel alive.
DATA: hard-code the routes (waypoint arrays), ports (lat/lng + TEU + rank), and chokepoints
(lat/lng + description + trade %) as JS objects. Make the dataset genuinely comprehensive.
STRUCTURE: projection + map render, route renderer (great-circle paths + animated flow),
ports layer, chokepoints layer, interaction/tooltips, UI dock + toggles + search, legend.
VERIFY: world map renders (flat + globe toggle, globe rotates); all major routes drawn as
curved animated lines; ports + chokepoints plotted + labeled; hover shows correct detail
cards; toggles show/hide layers; search works; legend accurate; smooth + premium looking.
Output the complete file, no truncation. This is a Data & Analytics showcase — the dataset
must be comprehensive and the map must look like a professional trade-intelligence tool.Round 4 winner: Fable 5. Score after four rounds: 2-2.
Round 5: 3D Hi-Fi Turntable Simulator
See the Pen Live Solar System - Fable by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
See the Pen Vinyl - HIFI Turntable Sim - KK3 by ਰਮਨਪਾਲ ਸਿੰਘ (@kwebby) on CodePen.
The final round asked for a realistic 3D vinyl turntable simulator: power button, tonearm, pitch and tone controls, volume, disk switching, and a cinematic auto-rotate camera mode. Both models delivered a working turntable with all controls functional, background art that changed with the selected record, and a cinematic view.
One build's tonearm and disk looked slightly off-model when the camera rotated wide. Both builds nailed the core interaction: power on, drop the needle, adjust pitch, switch vinyl.
Build Vinyl — a realistic 3D turntable + hi-fi music player SIMULATOR the user can
interact with in 3D. Not a flat music UI — a real 3D scene of a record player you operate.
DELIVER: self-contained vinyl.html. Three.js r128 via CDN + GSAP + Google Fonts allowed.
RUN AT: reasoning_effort max.
═══════════════ THE 3D SCENE ═══════════════
A photorealistic-leaning 3D turntable on a wooden surface in a warm-lit room, rendered
with Three.js WebGL. The user orbits, zooms, and interacts with real parts.
MODEL (build from Three.js primitives + custom geometry — no external model files):
- Turntable base (rounded box, matte/wood material)
- Platter (cylinder) with a spinning record (textured disc: grooves via canvas texture
or radial line shader, a center label with the current track art)
- Tonearm (articulated: pivot base + arm + headshell + stylus) that physically lifts
and swings onto the record when you press play, and lifts off on stop/eject
- Control knobs (rotary): volume, pitch/speed (33/45 RPM), tone
- Power button, start/stop, a cueing lever
- Optional: two speaker cabinets beside it with cones that subtly pulse to the music
MATERIALS & LIGHTING (this is the visual test):
- PBR-style MeshStandardMaterial: wood grain, matte plastic, brushed metal tonearm,
glossy vinyl with a subtle reflective sheen
- Warm key light + soft fill + a rim light; soft shadows (shadow maps)
- Environment reflection (a simple gradient env map / PMREM) for the metal + vinyl
- Subtle bloom on the power LED and highlights
═══════════════ INTERACTION & SIMULATION ═══════════════
- OrbitControls-style camera (implement manually if needed): drag to rotate, scroll to zoom
- Click the power button → LED on, subtle hum
- Click play / move the cueing lever → tonearm animates onto the record (GSAP),
record starts spinning (RPM depends on 33/45 selection), audio plays
- Drag the tonearm manually → scrub position on the track (real interaction)
- Turn the pitch knob → record spin speed + audio playback rate change together
(real turntable pitch-bend feel)
- Turn the volume knob → volume; tone knob → a lowpass filter cutoff
- Stop → tonearm lifts and returns, platter spins down with inertia
- Swap records: a small stack/crate of 3-4 albums beside the turntable; click one to
load it (record art + track changes, swap animation)
AUDIO:
- Use Web Audio API. Include 2-3 short royalty-free-style looping tracks generated
procedurally (synth pads/beats via oscillators) OR let the user drag-drop an audio
file onto the scene to play their own. (Default: procedural tracks so it works offline.)
- Playback rate ties to platter RPM + pitch knob (authentic turntable behavior)
- Volume + tone (lowpass) knobs affect the audio graph in real time
- A subtle vinyl-crackle layer for authenticity (synthesized noise)
- Speaker cones + a small spectrum visualizer react to the audio (AnalyserNode)
═══════════════ THE RICH 2D UI OVERLAY ═══════════════
A tasteful HUD layered over the 3D scene:
- Now-playing card (track title, album art, elapsed/total, progress)
- Transport controls (also operable via the 3D parts)
- A real-time audio spectrum / waveform visualizer
- RPM indicator (33/45), pitch % readout
- Record crate / playlist selector
- A "view" toggle: front / top / cinematic auto-orbit
- Warm, hi-fi aesthetic: cream/charcoal/brass palette, clean type
PERFORMANCE: cap pixel ratio at 2, dispose properly, keep 60fps. Reduced-motion fallback.
STRUCTURE: organize into scene setup, model builders (buildPlatter, buildTonearm,
buildKnobs...), interaction handlers, audio engine, UI overlay, animation loop.
VERIFY: camera orbits; power/play/stop work; tonearm animates onto/off the record;
record spins at correct RPM; pitch knob bends speed+audio together; volume/tone work;
swapping records changes art+track; visualizer reacts; looks genuinely like a hi-fi.
Output the complete file, no truncation. This must read as a REAL 3D turntable simulation,
not a flat player with a spinning circle.Both outputs looked stunning in their own way, and each had a small rough edge. I called this one a tie.
Round 5 winner: Tie. Final score: Kimi K3 3, Fable 5 3.
What Kimi K3 Actually Is
Kimi K3 is Moonshot AI's flagship model, released on July 16, 2026. Moonshot, a Beijing-based AI lab, built K3 as a 2.8-trillion-parameter mixture-of-experts model that is natively multimodal and ships with a 1,048,576-token context window (MarkTechPost, VentureBeat).
Two architectural pieces sit under the hood: Kimi Delta Attention, a hybrid linear-attention mechanism, and Attention Residuals, a drop-in replacement for standard residual connections that Moonshot says improves scaling. K3 also runs an always-on reasoning mode Moonshot calls "thinking mode," and the model weights use MXFP4 quantization-aware training, which keeps the file size (roughly 594 GB) from growing in proportion to the parameter count (MarkTechPost).
Moonshot is calling K3 the largest open-weight model released to date (VentureBeat). "Open-weight" is the operative phrase, though. Full public weights are not out yet.
Open Weights and License
Moonshot has scheduled the full open-weight release for July 27, 2026. As of this writing, the official Kimi K3 repository has not landed on Hugging Face or GitHub. Based on Moonshot's pattern with Kimi K2, K2.5, K2.6, and K2.7 Code, all released under a Modified MIT license with a single attribution clause, K3 is expected to ship under the same terms, but that is not confirmed until the actual repo and license file appear (Hugging Face blog via community post).
I would not make a license-dependent decision, like shipping K3 inside a commercial product, until that file is public.
Pricing
Moonshot prices Kimi K3 API access at $3 per million input tokens and $15 per million output tokens, with a discounted $0.30 per million tokens on cache-hit input. Pricing is flat across the full 1M-token context window, with no long-context surcharge (platform.moonshot.ai).
Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. That puts Kimi K3 at roughly one-third of Fable 5's list price on both input and output, not the one-fourth figure I threw out during the video. Close, but worth correcting for accuracy.
Benchmarks: How K3 Stacks Up Against Fable 5
Moonshot published a set of coding and reasoning benchmarks alongside the K3 launch. Independent trackers have compared these against Fable 5, Claude Opus 4.8, and GPT-5.6:
I'd treat every one of these numbers as directional, not gospel. Different labs run these evaluations through different agent harnesses: KimiCode for K3, Claude Code or Codex for the US models. Harness choice alone can swing a score by several points (BinaryVerse AI, Wan 2.7 blog).
My blind test agrees with the general shape of these numbers: Kimi K3 won on the builds that leaned on raw coding output and interactivity (the floor plan walkthrough and the racing game), while Fable 5 won on the rounds that came down to careful instruction-following (stopping the globe on command, cleaner state handling in the grand tour).
Demand and Availability
Moonshot paused new Kimi K3 subscriptions shortly after launch because user request volume in the first 48 hours pushed the company close to its GPU capacity limit. Existing subscribers keep their access; the company says it's adding capacity and will reopen signups in batches, alongside two new plan tiers: a Kimi Membership for the web app and a separate Kimi Code Membership for coding workflows (ANI News).
That pause matches what I ran into myself: when I tried to grab a Kimi K3 membership after recording this comparison, sign-ups were closed with a "new plans coming soon" message.
My Verdict
I called it a 3-3 tie, and I stand by that after digging into the official numbers. Kimi K3 is the one to reach for when you want raw interactive output on a budget. It's roughly a third of Fable 5's price, and it won the two rounds that came down to visual polish and interactivity. Fable 5 is the one to reach for when the build depends on the model doing exactly what you told it to do, especially on longer, multi-step instructions.
If you're picking a default for agentic coding work today, I'd run both on your actual workload before committing. The benchmark gap between them is real but small enough that your specific prompts and harness will matter more than the leaderboard number.
Frequently Asked Questions (FAQs)
Is Kimi K3 open source?
Kimi K3 is open-weight, meaning Moonshot plans to release the model weights for download and self-hosting. Full weights are scheduled for July 27, 2026, and the license is expected to follow the Modified MIT terms Moonshot used for its earlier K2 models, though this is not officially confirmed yet.
How much does Kimi K3 cost to use through the API?
Moonshot charges $3 per million input tokens and $15 per million output tokens for Kimi K3, with a $0.30 per million token rate on cached input. That makes it roughly one-third the cost of Claude Fable 5, which is priced at $10 input and $50 output per million tokens.
Is Kimi K3 better than Claude Fable 5?
It depends on the task. In my blind test, Kimi K3 won on raw interactive builds and graphics polish, while Fable 5 won on precise instruction-following. Independent benchmarks show K3 ahead on some coding tests (SWE Marathon) and behind on others (DeepSWE, HLE-Full), so neither model wins across the board.
Can I sign up for Kimi K3 right now?
New subscriptions were paused shortly after the July 16, 2026 launch because demand exceeded Moonshot's available GPU capacity. Existing subscribers keep access, and Moonshot says it will reopen signups in batches with new membership tiers.
What is Kimi Delta Attention?
Kimi Delta Attention is a hybrid linear-attention mechanism Moonshot built into Kimi K3's architecture, paired with a technique called Attention Residuals. Moonshot says the combination improves how the model scales compared to standard attention and residual designs.
Can I run Kimi K3 locally?
Not yet. Moonshot has not published the official model weights as of this writing. The company has scheduled the full open-weight release for July 27, 2026, at which point self-hosting guides will follow.
Final Thoughts
Five rounds, one blind test, and a dead-even 3-3 score: Kimi K3 and Claude Fable 5 are closer than the hype cycle around either one suggests. Kimi K3 backs up its price advantage with real output quality on interactive builds, and it's worth your attention if budget matters to your workflow. Run your own blind test with your own prompts before you switch your default model. That's the only benchmark that actually reflects what you build.
If you want the exact prompts from every round in this comparison, grab them from promptslove.com.





