OTHER April 13, 2026

Kimi K2.6 Released: 5x Cheaper Than Claude, 12-Hour Coding Sessions

WHAT CHANGED

Moonshot AI released Kimi K2.6 on April 13, 2026 — an open-weight model handling 12-hour coding sessions and 300 parallel agents. Pricing: $0.60/$2.50 per million input/output tokens vs Claude Sonnet 4.6's $3.00/$15.00 (5-6x cheaper). 4,000-step coordinated agent ceiling.

WHY IT MATTERS

First credible Western+Chinese parity moment for coding models. K2.6 trades a smaller context window (262K vs 1M) for radically lower cost and self-hostable weights. For cost-sensitive teams running coding agents at scale, this is a budget-level decision, not a marginal one.

HOW TO USE IT

Use kimi-cli 1.33.0 with the Kimi for Code subscription, or plug K2.5 weights into Claude Code by setting ANTHROPIC_BASE_URL to the Moonshot endpoint. K2.5 weights live on Hugging Face under Modified MIT license; K2.6 weights expected soon.

OTHER / BASH

# Plug Kimi into Claude Code as backend
export ANTHROPIC_BASE_URL=https://api.moonshot.cn/anthropic
export ANTHROPIC_API_KEY=your-moonshot-key
claude

kimimoonshotopen-sourcecodingagents

ORIGINAL SOURCE

https://www.buildfastwithai.com/blogs/kimi-code-k26-preview-2026

VIEW ORIGINAL SOURCE →

K2.6 hits 66.7% on Terminal-Bench 2.0 (Terminus-2 harness), 96.4% on AIME 2026, and 90.5% on GPQA-Diamond. The 4,000-step agent ceiling is Moonshot’s specific architectural claim — neither Anthropic nor OpenAI publishes equivalent runtime numbers. Independent verification is still pending.

← BACK TO UPDATES