Kimi K2.7 Code: Open-Weights That Undercuts Claude by 12×
Agentic AIThe most capable open-weights coding model yet just landed at a price that makes the closed-model math hard to defend for high-volume agentic workloads.
The Number
Moonshot AI’s Kimi K2.7 Code — a one-trillion-parameter mixture-of-experts model with 32 billion active parameters — prices input at $0.95/M tokens and output at $4.00/M tokens. Claude Fable 5 output runs at $48.00/M tokens. That is the source of the 12× gap that The Decoder’s analysis documents.
GPT-5.5 output prices at approximately $30.00/M tokens, leaving Kimi K2.7 roughly 7× cheaper there as well. For teams running agentic pipelines where output costs typically dominate total spend by a factor of 3–5× over input, these are not marginal differences.
How the Math Works
The pricing positions Kimi K2.7 Code as a high-throughput agent workhorse, not a frontier reasoning replacement. Performance benchmarks reflect that distinction: on Program Bench, K2.7 scores 53.6 against GPT-5.5’s 69.1 — a meaningful gap on raw coding capability. On MCPMark Verified, the agent-specific testing suite, K2.7 Code leads at 81.1, ahead of Opus 4.8’s 76.4 and behind GPT-5.5’s 92.9.
The model carries a 256,000-token context window and full multimodal support across text, images, and video. As an open-weights model it can be deployed locally or through any inference provider, removing API vendor lock-in from the cost calculation entirely.
What It Signals
The Decoder’s framing captures the practical implication: “the same budget lets you run it many times more often.” For agentic pipelines where a single user interaction can spawn 20–50 model calls, cost compounds quickly. Running K2.7 Code at 12× lower output cost means pipeline economics that were previously viable only at enterprise scale become accessible to smaller teams.
This is the pattern that drove the shift toward cheaper capable models in 2024–2025, now repeating one tier up. Token pricing has already emerged as a core business metric for teams operating agentic systems at scale — and this release makes the cost of using frontier models for non-frontier subagent tasks far more visible.
The So What
Kimi K2.7 Code is not a replacement for frontier reasoning on tasks that require it. It is a serious candidate for any agentic subagent role — code review passes, tool-use loops, document processing, test generation — where output volume is high and the task does not require state-of-the-art performance. Teams that have not built their agentic architectures around model-per-task cost modeling are paying frontier prices for non-frontier work. This release makes that tradeoff concrete.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →