Breaking

Moonshot AI's Kimi K3 Joins Arena.ai's Agent Arena for Long-Horizon Task Evaluation

July 16, 2026 at 13:07 EDT

Foundation Models
AI Agents

Moonshot AI's newest flagship model, Kimi K3, is now available for public testing on Arena.ai's Agent Arena, the community-driven platform that evaluates AI systems on long-running, real-world agentic tasks. The announcement, made on July 16, 2026, positions the Chinese lab's latest model against rivals in an evaluation setting built to measure how models perform on messy, multi-step work rather than static benchmarks.

July 16, 2026 · Moonshot AI

Kimi K3 Enters the Arena

Moonshot AI's new flagship model goes public on Agent Arena — a platform that scores models on long-running, real-world agentic tasks rather than static benchmarks. Its headline feature: a one-million-token context window.

Token context window, up from 256k

~2.5T

Reported MoE parameters (unconfirmed)

Arenas: Agent, Text, Vision, Document, Frontend Code

The context leap

K3 handles nearly 4× the context of its predecessors in a single pass.

256k

K2.6 / K2.7 Code

Kimi K3

Pricing per million tokens

More capability, higher cost. Output pricing nearly quadruples versus K2.

Cache hit

$0.30

Input

$3.00

K2: ~$0.95

Output

$15.00

K2: ~$4.00

Praise

Unusually strong outputs, especially frontend code
Single prompts yield 3D, frontend, and small games
Compared to top-tier rival systems

Caveats

Some found in-arena responses very slow
Others await full official release before judging
Parameter count remains unconfirmed

How it's judged

A shift toward sustained, tool-using performance over one-off benchmark scores.

Millions of long-horizon tasks

→

Web, filesystem & terminal tools

→

Community voting & causal tracing

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.