Moonshot Releases Kimi K2.7-Code — Open Weights at $0.95/$4 per MTok

Moonshot AI released Kimi K2.7-Code on June 12, 2026, announcing it via kimi.com/code and publishing the full model weights to Hugging Face the same day. The model ID is kimi-k2.7-code. It is a coding-first release: a 1-trillion-parameter Mixture-of-Experts model with roughly 32B active parameters across 384 experts, a 256K-token context window, and a Modified MIT license that allows self-hosting. For a market where the frontier-tier coding models are closed and priced at $5–$10 per million input tokens, an open-weight 1T model at under a dollar is a genuinely different cost structure.

The pricing is the headline. Moonshot lists API access at $0.95 per million input tokens and $4.00 per million output tokens. For comparison, that is roughly a fifth of Claude Opus 4.8's $5/$25 and under a tenth of the now-suspended Fable 5's $10/$50 — for a current cross-model price comparison see the AI Tool Pricing Database. The model pairs with Kimi Code, Moonshot's terminal-first coding agent, with membership plans listed from $19 per month. For high-volume agentic workloads — where output token counts dominate the bill — the gap is large enough to change which workloads are economically viable to automate.

Open weights change the deployment math. Because the full weights ship on Hugging Face under a Modified MIT license, K2.7-Code is not only an API — it is a model teams can run themselves. That opens self-hosting and air-gapped deployment for organizations that cannot send code to a third-party API at all: regulated industries, defense-adjacent work, or any team with a hard data-residency requirement. The closed frontier models cannot serve that segment regardless of price. See our roundup of the best AI coding tools for where self-hostable options fit.

Moonshot's efficiency claim. The company reports that K2.7-Code uses roughly 30% fewer "thinking" tokens than its predecessor K2.6 while scoring higher on Moonshot's coding benchmarks — a meaningful claim, because reasoning-token usage is a direct driver of both latency and cost on agentic tasks. If it holds up, the effective cost advantage is larger than the headline per-token price suggests.

The caveat operators cannot skip. Every benchmark Moonshot has published for K2.7-Code is self-run and self-reported — it reports +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite, all relative to K2.6, with the company running the evaluations itself. As of release, there are no independent third-party numbers for K2.7-Code on the standard public leaderboards — a gap reporters flagged at launch. Vendor-reported gains on a vendor's own benchmark are a starting point, not a verdict. The improvement-over-our-last-model framing also says nothing about how it compares to the closed frontier models most teams actually weigh it against.

What it means for builders. Kimi K2.7-Code is best read as a cost-and-control play rather than a confirmed capability leap. If you run high-volume agentic coding and your bill is dominated by output tokens, the price difference is worth a real evaluation. If you have data-residency constraints that rule out closed APIs, the open weights are the actual unlock. But until independent benchmarks land, treat the performance claims as unverified and run the model against your own repository and your own task suite before trusting it on anything that matters.