Kimi K2.7 Code: Moonshot AI launches 1T-code model vs GPT-5.5

Moonshot AI has released Kimi K2.7 Code, an open-weights, one-trillion-parameter model trained for programming tasks, and priced to compete on inference cost. The company positions the model as a low-cost option for code generation and related developer workflows, and initial comparisons show Kimi undercutting OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 on price per token by as much as 12x.

Kimi K2.7 Code is being distributed with weights available to the public, enabling teams to run the model on their own hardware or through third-party hosts. Moonshot AI highlights the model's focus on code completion, synthesis, and code-aware reasoning, and publishes example latency and throughput figures for common GPU configurations. While the model's parameter count is one trillion, Moonshot cautions that Kimi does not yet match proprietary models on a range of quality benchmarks.

Benchmarks and pricing

Early benchmark snapshots place Kimi behind GPT-5.5 and Claude Opus 4.8 on many standard code-evaluation metrics, including functional correctness on code synthesis tasks and multi-step reasoning in complex debugging scenarios. The gap is not uniform, with Kimi performing competitively on single-step completions and boilerplate generation, while more sophisticated reasoning and long-context problems favor the closed models.

Where Kimi stands out is cost. Moonshot's published pricing and community-hosted instances put Kimi's effective price per token significantly below the leading managed APIs. Aggregated comparisons show Kimi can be up to 12x cheaper per token than GPT-5.5 and Claude Opus 4.8 depending on deployment and throughput assumptions. That gap narrows if customers use highly optimized managed endpoints from established providers, but it remains material for teams running large volumes of code inference in-house.

The open-weights nature of Kimi lowers entry barriers for on-premise and private-cloud deployments, and it enables organizations to tailor inference stacks for cost or latency. Moonshot supplies recommended configurations and quantized checkpoints to reduce memory and compute needs, and community maintainers are already publishing optimized containers and orchestration tips.

Quality trade-offs, however, are real. On head-to-head functional tests, GPT-5.5 and Claude Opus 4.8 maintain higher pass@k and fewer hallucinated function calls. Moonshot acknowledges these gaps in its technical notes and frames Kimi as a pragmatic option when cost, control, or data sovereignty are priorities.

Why it matters

An open-weights, trillion-parameter code model priced far below leading closed models changes the calculus for teams that run heavy code inference workloads or require private deployments. Developers and smaller companies can now choose lower-cost, locally controlled models for many production tasks, while organizations that need the highest correctness and reasoning will still favor proprietary offerings. The release intensifies pressure on API providers to justify managed pricing with measurable quality or integration advantages.