Kimi K3 Is Second Only to Claude Fable 5 — But at $10/Task, Is It Worth It?

The model leaderboard just reshuffled again. Last week Moonshot AI released Kimi K3, a 2.8 trillion parameter model that now sits at #3 on the Artificial Analysis Intelligence Index (score: 57), rubbing shoulders with Opus 4.8 and GPT-5.5. On the newer, more agentic AA-Briefcase benchmark, Kimi K3 hits an Elo of 1543 — a staggering +727 improvement over Kimi K2.6 and second only to Claude Fable 5 (1574).

But the cost story is very different. Before you wire Kimi K3 into your agent pipeline, here's what the numbers actually say.

What Is AA-Briefcase?

AA-Briefcase is a proprietary benchmark from Artificial Analysis that tests models on realistic knowledge-work projects — generating spreadsheets, presentations, UI mockups, and analytical reports from thousands of linked input files. Performance is scored as a single Elo combining:

• Objective correctness (rubric pass rate)
• Analytical quality (reasoning depth, accuracy)
• Presentation quality (formatting, clarity)

It's designed to measure what engineers actually care about: can the model do the whole job, not just answer a multiple-choice question.

Where Kimi K3 Excels

Kimi K3's AA-Briefcase results reveal a model that's genuinely competitive at the frontier:

Metric	Kimi K3	Claude Fable 5	GPT-5.6 Sol (max)
Overall Elo	1543	1574	1501
Rubric pass rate	51%	56%	41.8%
Analytical quality Elo	1754	1744	—
Presentation quality Elo	1471	—	1660

The analytical quality score is the standout — 1754 Elo is within measurement noise of Fable 5's 1744. Kimi K3's reasoning is genuinely frontier-class.

The Fireworks AI team reported that Kimi K3 and Fable 5 achieve complementary SoTA — meaning they excel at different task types and can be used together in ensemble or routing setups for best results. This is a meaningful finding for teams running production inference pipelines.

Where It Falls Short: Cost and Latency

Here's the catch. Frontier performance doesn't come cheap — and Kimi K3 is expensive in both time and money.

Cost per task: $10.57 — placing it among the most expensive models on AA-Briefcase. Pricing is $3/$15 per 1M input/output tokens (with a 90% discount on cached tokens).

Time per task: 56.4 minutes on average. That's ~2.5x slower than Claude Fable 5 and ~3.8x slower than Grok 4.5 (high). The bottleneck is the sheer number of turns: Kimi K3 averages 83 turns per task (vs. 67 for Fable 5 and 50 for GPT-5.6 Sol), consuming 120k output tokens per task.

For context, Kimi K2.6 used 42k output tokens and 54 turns per task. Kimi K3 is generating nearly 3x more output to achieve its frontier-level scores.

Why This Matters Right Now

Kimi K3's release is part of a broader pattern. As the Artificial Analysis team noted, six labs now field a model scoring above 50 on the Intelligence Index — up from two in early June 2026. Grok 4.5, GPT-5.6, Muse Spark 1.1, and Kimi K3 all landed within an 8-day window.

The market is fragmenting. No single model dominates across cost, latency, and quality. Teams that treat model selection as a routing problem (cheap/latency-sensitive → Grok 4.5, analytical depth → Kimi K3, best overall → Fable 5) will beat teams that bet on one provider.

How to Think About Kimi K3

Use it when: You need deep analytical reasoning on complex, long-horizon tasks where accuracy matters more than cost or speed. The analytical quality is genuinely frontier-class.

Avoid it when: You're building real-time or high-volume pipelines. 56 minutes and $10.57 per task won't work for chat, code completion, or anything with a sub-second SLA.

The ensemble play: If you have the infrastructure, routing between Kimi K3 (analysis) and Fable 5 (generation) appears to give best-of-breed results — the "Kimi K3 + Fable = SoTA" finding from Fireworks is worth taking seriously.

Bottom line: Kimi K3 proves that the frontier isn't an OpenAI-Anthropic duopoly anymore. But raw capability isn't the same as practical utility. Benchmark scores are rising faster than cost-efficiency, and that gap is the real engineering problem of 2026.

---

Sources: Artificial Analysis — Kimi K3 on AA-Briefcase, Fireworks AI — Kimi K3 vs Fable, Artificial Analysis — Four Frontier Launches in Eight Days.