Kimi K3 Is Fable Level... (they should be worried)

Summary
In this video, a presenter from Better Stack reviews Moonshot AI’s newly announced Kimi K3 model, examining its technical specifications, benchmark performance, and API pricing against models from Anthropic and OpenAI. He tests Kimi K3 alongside Claude Fable 5, GPT-5.6 Sol, and GLM-5.2 on two real-world coding challenges: building a browser-based 3D Formula 1 racing game and developing a full-stack personal finance dashboard.

What is shown

• [00:18] Overview of the official announcement post, “Kimi K3: Open Frontier Intelligence”, detailing architecture and internal benchmarks across coding and agent evaluations.
• [01:02] Artificial Analysis benchmark charts comparing Kimi K3’s Coding Index, Agentic Index, and Intelligence Index scores against competing models.
• [01:52] Comparison of API pricing and cost/time per task on the Artificial Analysis Intelligence Index.
• [02:42] Test 1 (Three.js F1 Racing Game): Kimi K3 (via OpenRouter and Kimi Code CLI), Claude Fable 5 (via Claude Code), GPT-5.6 Sol (via Codex), and GLM-5.2 are tasked with generating a browser-based 3D racing game. The presenter plays through each generated game to evaluate rendering, physics, controls, and lap logic.
• [05:23] Test 2 (Personal Finance Dashboard): Prompting Kimi K3, Fable 5, GPT-5.6 Sol, and GLM-5.2 to generate a full-stack personal finance dashboard application, inspecting the UI, interactive charts, and underlying code architectures (React, Express, SQLite, Next.js, and Drizzle ORM).
• [09:50] Highlights from Moonshot AI’s launch cases, including an open-world procedural 3D browser exploration game and a chip design case study where Kimi K3 designed a nano-model accelerator chip.
• [10:27] Announcement screen stating that full model weights are scheduled for release by July 27, 2026.

Claims & numbers

• The presenter notes that Kimi K3 is a 2.8-trillion parameter model built with Kimi Delta Attention and Attention Residuals, featuring native vision and a 1-million-token context window.
• On Artificial Analysis leaderboards shown:
  • Coding Index: Kimi K3 scores 76.2 (0.3 points behind Claude Fable 5 at 76.5; GPT-5.6 Sol leads at 77.4).
  • Agentic Index: Kimi K3 scores 50.1, placing 3rd behind GPT-5.6 Sol (54.0) and Claude Fable 5 (52.8).
  • Intelligence Index: Kimi K3 scores 57, behind GPT-5.6 Sol (60) and Claude Fable 5 (59).
  • DeepSWE: Kimi K3 reaches 67.5% on Moonshot’s benchmark chart (behind GPT-5.6 Sol at 73.0% and Fable 5 at 70.0%).
• Pricing: Kimi K3 API pricing is listed at $3 per million input tokens and $15 per million output tokens, which the presenter notes matches Claude Sonnet 5 and is cheaper than Claude Fable 5 ($10/$50), Claude Opus 4.8 ($5/$25), and GPT-5.6 Sol ($5/$30).
• On the Artificial Analysis weighted cost per Intelligence Index task, Kimi K3 costs $0.94, compared to $0.47 for GLM-5.2, $1.04 for GPT-5.6 Sol, $1.83 for Claude Opus 4.8, and $2.11 for Claude Fable 5. Average decode time per task is 6.0 minutes.
• In the local Three.js F1 test, Kimi K3 spent 14 minutes 55 seconds thinking, took 35 minutes 20 seconds total to build the app, and cost $1.24 in tokens; Fable 5 took 44 minutes 36 seconds; GPT-5.6 Sol took 18 minutes 15 seconds; GLM-5.2 took 5 minutes 50 seconds.
• In the finance dashboard test, Kimi K3 via OpenRouter took 56 minutes 29 seconds ($1.30 cost); Fable 5 took 56 minutes 35 seconds; GPT-5.6 Sol took 31 minutes 39 seconds; GLM-5.2 took 15 minutes 47 seconds ($1.11 cost).
• In the chip design case study, the presenter reports that Kimi K3 completed an autonomous 48-hour run to design, optimize, and verify an accelerator chip using Nangate 45nm EDA tools.
• The presenter highlights Moonshot’s commitment to release the full model weights by July 27, 2026.

Notable quotes

• [00:03] “It didn’t just catch up with Opus and GPT-5.5, it caught up with Fable and GPT-5.6.”
• [01:42] “Like, this model just seems too good to be true.”
• [09:26] “So from my quick tests there, I think Kimi K3 does stand up with Fable and GPT-5.6 Sol, and live up to the hype that the benchmarks are giving it.”

Assessment
This is an independent hands-on review and comparison demonstrating real coding outputs generated by Kimi K3 and rival frontier models. The local coding tests and UI walk-throughs are authentic and transparent, noting minor glitches (such as inverted steering controls in Three.js and unstyled track textures) while validating the model’s high benchmark rankings.

Described by gemini-3.8-flash on 2026-10-12 from the video’s audio and frames.