Breaking

Moonshot AI's Kimi K3 Can Now Run Inside OpenAI's Codex via a Third-Party Router

July 19, 2026 at 21:15 EDT

Software Dev & Coding
Foundation Models
AI Agents

Developers have begun wiring Moonshot AI's new Kimi K3 model into OpenAI's Codex agentic coding environment, using an open-source router that reportedly makes the setup take about five minutes. The workaround lets users drive Codex's command-line and app interface with a rival Chinese frontier model rather than OpenAI's own system.

Agentic Coding · Model × Harness

Kimi K3 slots into OpenAI's Codex — in about 5 minutes

Developers are wiring Moonshot AI's new flagship into Codex through an open-source router, driving a rival Chinese frontier model inside OpenAI's own command-line interface with minimal reconfiguration.

The ~5-minute wiring

CODEX CLI

Familiar OpenAI workflow

→

CC SWITCH

Open-source router / translator

→

KIMI K3 API

OpenAI-compatible endpoint

2.8T

parameters, Mixture-of-Experts

1.05M

token context window

~5 min

to complete the setup

Published Kimi K3 scores

Via Moonshot's KimiCode harness at max reasoning effort — results shift by harness.

88.3

Terminal-
Bench 2.1

77.8

Program
Bench

42.0

SWE
Marathon

What developers praised

Strong on design-oriented and long-horizon coding tasks
Handled work more capably than Claude in some cases
Completed security fixes other assistants refused
Felt stronger than Codex's own tuned system in the same UI

Practical caveats

Codex harness may not be the optimal environment
Heavier token use raises latency and cost
Reasoning effort tends to lock at "max"
Full checkpoint reportedly due July 27, 2026

The takeaway

The harness and the model are becoming separable choices — frontier models from competing labs now swap into rival tooling through lightweight routers.

Continue reading

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