Kimi K3.1 Pro API

Kimi K3.1 Pro is a multimodal large language model developed by Moonshot AI and released on October 23, 2026. As the first model in the K3.1 series and the most powerful 3-trillion-parameter model to date, it delivers significant leaps in long-horizon project programming, parametric 3D creation, end-to-end financial research, and complex long-running agent workflows. It reduces the complex reasoning hallucination rate from K3's 23.4% down to 8.3%.

On the iCreat platform, Kimi K3.1 Pro is accessible via an OpenAI-compatible Chat Completions endpoint at https://api.icreat.ai/v1. It supports a massive 1,048,576-token context window and native multimodal input (text, image, and video). The pricing is aligned with Kimi's official overseas rates: $2 per million tokens for input, $10 per million tokens for output, and $0.2 per million tokens for cache reads.

Model Positioning

Kimi K3.1 Pro is positioned as a flagship open-weights multimodal large language model tailored for complex, long-horizon agentic workflows and advanced coding tasks. By leveraging a 2.8T total parameter Mixture-of-Experts (MoE) architecture with 104B active parameters, it balances deep reasoning capabilities with efficient inference. It is designed for developers and enterprises requiring top-tier performance in software engineering, financial modeling, and 3D creation, offering a 1-million-token context window and native vision understanding at a highly competitive price point compared to peer flagship models.

Core Capabilities

Long-Horizon Project Programming

Kimi K3.1 Pro excels in sustained, complex coding tasks. It achieves 51.7% on the SWE-Marathon benchmark and 61.2% on the Kimi Code Bench 3.0 Long-horizon, leading all models in long-context software engineering and agentic coding workflows.

Parametric 3D Creation and Modeling

The model brings significant improvements to parametric 3D creation and modeling, enabling developers to generate and manipulate complex 3D structures through natural language instructions seamlessly.

End-to-End Financial Research

With a score of 64.4 on OfficeQA Pro v2 (claimed SOTA), Kimi K3.1 Pro can ingest and analyze massive financial documents, facilitating end-to-end financial research and modeling within its 1-million-token context window.

Complex Long-Running Agent Workflows

Achieving 65.7% on Terminal-Bench 4.0, the model is highly capable of navigating complex, multi-step agent workflows, tool usage, and autonomous task completion.

Reduced Hallucination in Complex Reasoning

Through architectural and training refinements, the hallucination rate on the Kimi Hallu Bench Hard has been reduced by approximately 65%, dropping from 23.4% in the previous generation to 8.3%.

Native Multimodal Understanding

The architecture integrates the MoonViT-V2 vision module (401M parameters), providing native understanding of text, images, and video inputs without relying on external adapters.

Pricing

Note: The above is the iCreat platform price, consistent with Kimi's official overseas pricing (K3.1 Pro input $2, output $10, cache read $0.2 per million tokens). This is approximately 33% cheaper than the previous K3 pricing ($3/$15). iCreat platform pricing prevails.

Application Scenarios

• Complex long-horizon software engineering and autonomous coding agents
• End-to-end financial document analysis, research, and quantitative modeling
• Parametric 3D content creation and structural modeling
• Multi-step autonomous task execution and tool-use workflows
• Processing and reasoning over massive documents (up to 1 million tokens)
• Multimodal data analysis (combining text, images, and video)

Model Comparison

Comparison Table 1: Kimi K3.1 Pro vs Kimi K3

Note: K3.1 Pro is approximately 33% cheaper per token than K3. According to third-party Artificial Analysis, the single-task cost is $2.11 for K3.1 Pro vs $2.00 for K3 (K3.1 Pro outputs 14% more tokens), making the single-task cost comparable while delivering significantly stronger capabilities.

Comparison Table 2: Cross-Vendor Flagship Comparison

Note: K3.1 Pro pricing matches Gemini 4 Argon's promotional rate and is approximately half the price of Claude Opus 5.5. Its coding agent capabilities approach those of Opus 5.5, while Argon leads on DeepSWE v1.1 (77.9%).

Why Choose Kimi K3.1 Pro?

• Unmatched long-horizon coding capability, leading the SWE-Marathon benchmark with a score of 51.7
• Massive 1-million-token context window for processing extensive documents and complex agent workflows
• Significantly reduced hallucination rate (8.3%) ensuring reliable reasoning in critical tasks
• Native multimodal understanding for text, image, and video inputs
• Highly competitive pricing at $2/$10 per million tokens, offering flagship performance at roughly half the cost of comparable models
• Open-weights availability planned for November 2026 under the Apache 2.0 license

Specifications

Architecture

Kimi K3.1 Pro is built on a 2.8T total parameter Mixture-of-Experts (MoE) architecture, activating 104B parameters per token. It utilizes the Stable LatentMoE framework, selecting 16 routed experts out of 896, plus 2 shared experts for each token. The model employs Kimi Delta Attention (KDA) hybrid linear attention combined with Attention Residuals (AttnRes). It features a 93-layer mixed stacking architecture, interleaving 69 KDA layers and 24 Multi-Head Latent Attention (MLA) layers in a 3:1 ratio. Native vision understanding is powered by the integrated MoonViT-V2 module, which comprises 401M parameters.

Notes

• The model parameter must use the iCreat model code: kimi-k3.1-pro.
• iCreat pricing ($2/$10/$0.2) is consistent with Kimi's official overseas pricing. Actual settlement is subject to platform billing and the costUSD field.
• K3.1 Pro enables reasoning (thinking) by default. The reasoning_effort parameter can be set to low, high, or max (default: max). For simpler tasks, lowering the reasoning effort can balance quality, latency, and token usage.
• The context window supports up to 1,048,576 tokens (1 million).
• The model natively supports visual understanding (text, image, and video inputs).
• Open weights are planned for release in November 2026. A K3.1 Pro Highspeed variant is also available, offering 2–3x faster inference speeds.

Frequently Asked Questions

How do I call Kimi K3.1 Pro?

You can call the model using the OpenAI-compatible /v1/chat/completions endpoint. Set the model field in the request body to kimi-k3.1-pro, use Bearer token authentication, and you can enable streaming by setting stream=true.

How is the cost calculated?

The cost is calculated based on token usage: $2 per million input tokens, $10 per million output tokens, and $0.2 per million cache read tokens. This is consistent with the official overseas pricing and is approximately 33% cheaper than the previous K3 generation. The usage field in the response provides the token statistics.

How large is the context window?

The context window is 1,048,576 tokens, which equals 1 million tokens.

How do I configure the reasoning effort?

The reasoning_effort parameter offers three levels: low, high, and max (default is max). For simpler tasks, lowering the reasoning effort can help balance generation quality, latency, and token consumption.

What is the difference between Kimi K3.1 Pro and Kimi K3?

K3.1 Pro is approximately 33% cheaper per token. Performance has improved significantly: Terminal-Bench 4.0 increased from 13% to 65.7%, and the hallucination rate dropped from 23.4% to 8.3%. Additionally, K3.1 Pro includes native vision understanding.

How does it compare to Claude Opus 5.5 and Gemini 4 Argon?

K3.1 Pro is priced at roughly half of Opus 5.5. On Terminal-Bench 4.0, it scores 65.7%, approaching Opus 5.5's 66.4%, and it leads on SWE-Marathon with a score of 51.7. Gemini 4 Argon leads on DeepSWE v1.1 (77.9%) and offers the same promotional pricing of $2/$10.