The Kimi K3 API gives developers hosted access to Moonshot AI's flagship model for long-horizon coding, multimodal analysis, tool use, and end-to-end knowledge work. Released on July 16, 2026, K3 is a 2.8-trillion-parameter Mixture-of-Experts model. It activates 16 of 896 experts per token and combines Kimi Delta Attention, Attention Residuals, and Stable LatentMoE to handle long sequences more efficiently than a dense model of similar total size.
K3 accepts text, image, and video input and returns text. Its 1M-token context window is intended for workloads that cannot be reduced to a short prompt: repository-scale coding, multi-document analysis, long tool histories, and visual feedback loops. The model always reasons. Developers can set reasoning_effort to low, high, or max, with max as the default, instead of switching thinking on and off.
GPTProto adds a different access layer around the model. One Kimi K3 API key and balance can also be used across 200+ text, image, video, and audio models. The rate displayed on this page is $2.70 per 1M input tokens and $13.50 per 1M output tokens, 10% below Moonshot's current 3/15 list price. This makes K3 easier to test beside GPT-5.6 Sol, Claude Fable 5, GLM-5.2, or other models without opening and funding a separate provider account for every experiment.
| Specification | Kimi K3 |
|---|---|
| Provider | Moonshot AI |
| Release date | July 16, 2026 |
| Model class | 2.8T-parameter Mixture-of-Experts |
| Active experts | 16 of 896 per token |
| Architecture | Kimi Delta Attention, Attention Residuals, Stable LatentMoE |
| Input | Text, image, and video |
| Output | Text |
| Context window | 1,048,576 tokens |
| Completion limit | 131,072 tokens by default; configurable up to 1,048,576 within the total context budget |
| Reasoning | Always on; low, high, or max effort |
| API features | Streaming, tool calls, tool_choice, dynamic tools, JSON mode, strict JSON Schema, and Partial Mode |
Kimi K3 API Applications
K3 is most useful when a task combines a large working set with repeated actions. A coding agent can inspect a repository, edit multiple files, run tests, read logs, and use screenshots to refine a frontend or game. A research workflow can keep a large document corpus in context, call external tools, and return a result that follows a strict JSON Schema. A multimodal QA system can compare interface screenshots with implementation details, while a video workflow can analyze uploaded footage and return scene notes, summaries, or structured metadata.
The model is less compelling for short classification, simple chat, or high-volume extraction where a smaller model already meets the quality threshold. K3 always reasons, so even its lower effort setting should be evaluated for latency and output-token cost. Start with one representative workflow, measure task completion rather than a single response, and compare K3 with a smaller model before routing all traffic to it.
API Details to Check Before You Migrate
K3 works with the OpenAI SDK and the Chat Completions request shape, but it has model-specific behavior that a simple model-name swap can miss.
First, Preserved Thinking is always enabled. In multi-turn conversations and tool loops, return the complete assistant message from the previous response, including reasoning_content and tool_calls. Keeping only the visible content can break reasoning continuity and make long sessions unstable. Historical reasoning also consumes context and is billed as token usage, so compact completed work instead of carrying every turn forever.
Second, use the top-level reasoning_effort field rather than the older K2.x thinking configuration. K3 supports low, high, and max. Its sampling values are fixed, so applications should not depend on custom temperature or penalty settings.
Third, parse structured results from the final content field, not from reasoning_content. K3 supports JSON mode and strict JSON Schema through response_format, which is useful for extraction pipelines, tool arguments, and machine-readable research outputs.
Finally, multimodal requests need structured content parts. Moonshot's native interface accepts base64 images or uploaded file references rather than public image URLs. Confirm the current GPTProto documentation for the exact image and video fields exposed by this endpoint before shipping a multimodal workflow.
When Should You Choose Kimi K3?
Moonshot's own launch post says K3 still trails Claude Fable 5 and GPT-5.6 Sol in overall performance. The reason to choose K3 is not a claim that it wins every benchmark. Its strongest case is the combination of a 1M context window, native video understanding, long-horizon coding, and a substantially lower token price.
| Decision factor | Kimi K3 | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|---|
| Official API list price per 1M tokens | $3 input / $15 output | $5 input / $30 output | $10 input / $50 output |
| Context window | 1,048,576 | 1,050,000 | 1,000,000 |
| Max output | Up to 1,048,576 within context; 131,072 default | 128,000 | 128,000 |
| Native inputs | Text, image, video | Text, image | Text, image |
| Model access | Hosted API; full weights announced for July 27, 2026 | Closed API | Closed API |
| Best fit | Cost-aware coding, multimodal agents, large working sets | Highest-end coding, computer use, and OpenAI tool workflows | Long-running agents, complex vision, and premium knowledge work |
Choose K3 when its lower cost, video input, or open-weight roadmap matters more than winning the broadest set of frontier evaluations. Choose GPT-5.6 Sol or Claude Fable 5 when your own task-level evaluation shows that the higher completion rate offsets their higher token price. For a benchmark-by-benchmark discussion, read What Is Kimi K3—and Is It Really Close to GPT-5.6 and Fable 5?.
