Kimi K3 vs Fable 5, GPT-5.6 & Opus 4.8: We Tested the Hype

Kimi K3 is Moonshot AI's new flagship, and for a few days in July 2026 it was the most-tested model on the internet. It surfaced first on the Arena eval site as a mystery model codenamed "Kivine" — self-described as "from Moonshot labs" with a January 2025 knowledge cutoff — before Moonshot officially announced it on July 16, 2026, with open weights promised before July 27. The recurring headline from testers: Kimi K3 lands somewhere around Claude Fable 5, and often ahead of GPT-5.6 on creative and 3D work. This first-impressions review pulls the specs, the benchmark spine, and a dozen community tests into one place.

A note for builders: these are early, uncontrolled community tests, so treat them as directional, not scores. OrcaRouter fronts API-available models behind one OpenAI-compatible endpoint, so once Kimi K3's API is live you can trial it against Fable 5 and GPT-5.6 without wiring up multiple SDKs.

TL;DR verdict. Kimi K3 is the first open-weight model to credibly reach Fable 5 range on frontend, 3D, and creative one-shots — it tops the Arena frontend leaderboard and sits #4 on the Artificial Analysis Index. But it's verbose and slow (one frontend build took 35 minutes), trails Fable on agentic coding, and the "won" verdicts are early Arena impressions, not audited benchmarks. Impressive, not settled.

Key takeaways

•  Kimi K3 is a 2.8T-parameter MoE, positioned as the largest open-weight model to date, with open weights due before July 27.

•  On neutral third-party evals it ranks #4 of 189 on the AA Intelligence Index (57) and #1 on Arena's frontend-code board (1679), above Fable 5's 1631.

•  Priced at $3 / $15 per 1M tokens — Moonshot's most expensive model ever, but still roughly a third of Fable 5's $10 / $50.

•  Testers repeatedly call it "Fable level"; skeptics say the head-to-heads are cherry-picked and it still trails on multi-step agentic tasks.

•  Watch the caveats: verbose output (~2x peer-median tokens), ~34s first-token latency, and it's still behind Fable on FrontierSWE and DeepSWE.

The specs and benchmark spine

Here's the hard data, each figure attributed to its source.

•  Architecture — Kimi K3: 2.8T MoE (896 experts, 16 active), Kimi Delta Attention, 1M context, native vision; Comparison point: Largest open-weight model to date (Moonshot, vendor-reported)

•  Variants — Kimi K3: K3 Max (chat/agent), K3 Swarm Max (parallel); Comparison point: —

•  Pricing (in / out) — Kimi K3: $3 / $15 per 1M; cache-hit ~$0.30; Comparison point: Fable 5 at $10 / $50 (~3.3x more)

•  AA Intelligence Index — Kimi K3: 57 — #4 of 189; Comparison point: Behind two Fable 5 configs + GPT-5.6 Sol; above Opus 4.8, Sonnet 5, GLM-5.2 (Artificial Analysis)

•  Arena frontend-code — Kimi K3: #1 at 1679, won 6 of 7 categories; Comparison point: Above Fable 5's 1631; up from K2.6's #18 (Arena / LMArena)

•  GDPval-AA v2 / AA-Briefcase — Kimi K3: 1687 (#3) / 1527 (#2); Comparison point: Artificial Analysis

•  GPQA Diamond — Kimi K3: 93.5% (strongest open model); Comparison point: Vendor-reported

•  BrowseComp / HLE w/ tools — Kimi K3: 91.2% / 56.0%; Comparison point: Vendor-reported

•  Agentic SWE — Kimi K3: FrontierSWE 81.2, DeepSWE 67.5; Comparison point: Behind Fable 5's 86.6 / 70.0

•  Terminal-Bench 2.1 — Kimi K3: 88.3; Comparison point: Above Fable 5's 84.6 (harness-mixing caveat below)

Two things temper the wins. First, the price advantage is partly eaten by verbosity — K3 emits about twice the peer-median output tokens on comparable tasks — plus a ~34-second first-token latency. Second, the Terminal-Bench edge comes with a warning from @ChrissGPT: Moonshot's benchmarks "often like to use Terminal 2 instead of 2.1," so treat cross-harness comparisons carefully. Independently, before the official index dropped, @teortaxesTex estimated "~55ish AA… around Opus 4.8 or GPT 5.5."

Community test roundup

The evidence that made Kimi K3 a story is the flood of Arena one-shots. Here's a representative slice — tester, task, and verdict, each linked to the original post.

•  @synthwavedd — Task: General stress-testing; Verdict: "Another DeepSeek R1 moment… often Fable level, maybe a little worse, but consistently better than 5.6. A beast."

•  @chetaslua — Task: Coding, head-to-head vs Fable; Verdict: "Mogs every open-source model… on par with Fable, sometimes better quality"

•  @Gc_qube — Task: 3D / games vs Fable 5; Verdict: "Kimi K3 won… the first model that has caught up with Fable"

•  @abhinavflac — Task: Sakura bonsai render; Verdict: "Beat Claude Fable — nailed the twisted trunk and layered canopy"

•  @testingcatalog — Task: Universe-sim vs Fable 5; Verdict: "Fable finished faster with more robust UX; K3 much more complex and visually appealing… very close"

•  @israfill — Task: Same universe-sim prompt; Verdict: "K3 wins on dense creative output and 'how did it build that' moments… surprised how close this is"

•  @mirochill — Task: Frontend one-shot (French); Verdict: "Equals 5.6 Pro, better than Fable 5, in a single attempt" (lighting issues noted)

•  @jun_song — Task: Flappy Bird vs Opus 4.8; Verdict: "Significantly better than Opus… Opus-5 level"

•  @JustinGorya — Task: Single-file HTML Minecraft; Verdict: "A new milestone… incredible at 3D and front end" (reply: "This is one shot. Crazyyyyy")

•  @redkendl — Task: 3D paper-plane game; Verdict: "One-shotted… Fable 5 / GPT-5.6-level output. Chinese labs are not 8 months behind anymore"

•  @noctus91 — Task: Interactive 3D experiences; Verdict: "The level of detail, polish, and overall quality is honestly wild"

•  @Lentils80 — Task: Frontend page; Verdict: "VERY slow — took 35 minutes — but one of the best outputs I've ever seen from this prompt"

Cheerleaders vs skeptics

The bull case is straightforward: an open-weight model matching a closed frontier flagship on creative and frontend work, at a third of the price, is genuinely new. @notjazii called an early output "a new SOTA model dropping today, moonshot cooked," and @chetaslua framed it as a "deepseek moment again for OSS."

The skeptics push back hard, and their objections are the same three every time:

•  Cherry-picked demos. On the bonsai test, @Heeseon replied "anyone can see Fable 5 won," and @sebuzdugan noted "one bonsai sample shows style control, not reliability across seeds."

•  Fable is just better. On Gc_qube's thread, @victor_vibing said flatly: "Fable is much better. Incomparably so."

•  Same aesthetic, not real range. @_everythingism argued the outputs "all have the 'Claude aesthetic'… generating a specific kind of design again and again," and on Opus tests @MCharles10581 simply said "it seems like opus is better."

The most useful frame comes from @teortaxesTex, who hadn't even tested K3 himself. His point is about an effort/taste gap, not a capability wall: "even Fable and Sol make so much SLOP… Kimi doesn't so much fail at stuff as it makes less interesting and detailed stuff, doesn't go the extra mile." For human-in-the-loop agentic coding on objective tasks, he expects K3 "should come unreasonably close." And the strategic kicker: "every Chinese model that merely *maintains* the gap… is a heroic feat." That's the honest read — K3 closes the distance on measurable tasks and loses a little on taste, which is exactly what its benchmarks show.

FAQ

Is Kimi K3 as good as Fable 5?

On frontend and 3D one-shots, testers say yes — it tops Arena's frontend board (1679 vs 1631) and multiple head-to-heads called it a win or a tie. On agentic coding it still trails (FrontierSWE 81.2 vs 86.6), and skeptics argue Fable holds an edge on polish and reliability across seeds.

Is Kimi K3 open source?

It's open-weight: Moonshot promised to release the weights before July 27, 2026. At launch it was API/web only. As the largest open-weight model to date (2.8T MoE), that release is the main event.

How much does Kimi K3 cost?

$3 / $15 per 1M input/output tokens, with cache hits around $0.30. That's Moonshot's most expensive model ever (K2.6 was $0.95/$4), but still about a third of Fable 5's $10 / $50 — though verbose output narrows the real-world gap.

Is Kimi K3 better than GPT-5.6?

Testers lean yes on creative and 3D work — @synthwavedd called it "consistently better than 5.6" and @mirochill put it level with 5.6 Pro. But on the AA Index, GPT-5.6 Sol still ranks above K3, and Fable 5 and 5.6 Sol lead on Terminal 2.1.

When was Kimi K3 released?

Officially announced July 16, 2026, after surfacing on Arena as "Kivine" around July 14. Open weights were promised before July 27, 2026.

Bottom line

Kimi K3 is the strongest first impression an open-weight model has made since the original DeepSeek moment: Fable-5-range creative and frontend output, top of the Arena frontend board, at roughly a third of Fable's price. Just hold the enthusiasm at "directional" — it's verbose, slow, still behind on agentic coding, and every "won" above is an early Arena impression, not an audited benchmark. When the weights drop before July 27, the controlled tests will tell us whether the hype held.