Kimi K3.1 Pro Review: Moonshot AI 1M Context Model, API Pricing and Highspeed Variant

Kimi K3.1 Pro reached public availability on October 23, 2026 through Kimi.ai and the Moonshot API. The model delivers a native 1,048,576-token context window, multimodal input, and always-on reasoning at three effort levels. It targets coding agents, long-horizon tasks, CAD work, and office automation while posting lower per-token prices than its July 2026 predecessor.

Kimi K3.1 Pro Overview and Key Features

Moonshot AI built Kimi K3.1 Pro as a post-trained upgrade to the earlier K3 release. The architecture uses a mixture-of-experts design with 2.8 trillion total parameters and 104 billion active parameters per token. Input modalities include text, images, and video; output remains text only. The model supports tool calling, structured JSON output, and cacheable context with a five-minute or one-hour TTL option.

Two API endpoints launched simultaneously: kimi-k3.1-pro and kimi-k3.1-pro-highspeed. The highspeed variant keeps identical capabilities but targets two-to-three times faster output. Zero Data Retention is available for API users, while enterprise features remain off by default. Open weights are scheduled for release by November 4, 2026 after Moonshot removed cyber-related training data.

Registration occurs through the Kimi.ai platform or direct API keys. No desktop client or local runtime is offered yet. The interface exposes effort-level controls and caching toggles directly in the playground and SDKs.

Pricing, Benchmarks, and Competitor Comparison

Standard pricing on the Moonshot API stands at $2 per million input tokens, $10 per million output tokens, and $0.20 for cached input. Cache writes cost $2 for the short TTL and $4 for the longer TTL. The highspeed tier doubles those rates to $4 input and $20 output. No additional long-context fees apply, and the listed prices sit roughly 33 percent below Kimi K3 rates.

Vendor-reported scores at maximum effort include 65.7 percent on Terminal-Bench 4.0, 51.7 percent on SWE-Marathon, 64.4 percent on OfficeQA Pro v2, and 92.7 percent on BenchCAD. Artificial Analysis recorded a 67.2 percent Terminal-Bench result and placed the model fourth on its Intelligence Index with a score of 55.1. Output speed averages 49.7–52 tokens per second.

Independent user feedback remains limited two weeks after launch. Early forum posts note strong performance on multi-file code tasks and reduced hallucination rates compared with K3, while some developers report slower-than-expected output on the standard tier. Complaints center on the current lack of open weights and occasional tool-calling instability on complex agent workflows.

Limitations and Final Assessment

Kimi K3.1 Pro shows relative weakness on science and terminal-bench science subsets, scoring 38.1 percent in vendor tests. The model has not yet received broad independent evaluations, and open-weight licensing details remain pending. Output speed ranks near the bottom third of tracked models, which may affect real-time agent use cases even on the highspeed tier.

The combination of 1M native context, competitive pricing, and documented gains on coding and office benchmarks gives Kimi K3.1 Pro a distinct position among frontier models. Its value proposition depends on workload fit and tolerance for the current speed profile until open weights arrive in early November.