Kimi K2.7 Code: Moonshot AI's New Coding-Focused Agentic Model

What You'll Learn

• What Kimi K2.7 Code supports in the official API documentation
• How cache-hit, cache-miss and output token pricing work
• How the standard model differs from K2.7 Code HighSpeed
• How to test a coding agent before production deployment

Kimi K2.7 Code: the verified model profile

Kimi's official API documentation describes Kimi K2.7 Code as a dedicated coding model. The pricing page says it supports text, image and video input, thinking mode, dialogue and agent tasks. It lists a 256K context length and support for long thinking and deep reasoning.

The same documentation lists automatic context caching, ToolCalls, JSON Mode and Partial Mode. These features matter because a coding agent does more than return plain text. It may inspect files, call tools, produce structured output and continue work across a long context. See our AI model pricing comparison for broader token-cost context.

What the 256K context length means

A 256K context length is a capacity limit for the material available to a request. It does not mean that every coding task should send an entire repository. Large prompts can raise cost, increase latency and make it harder to identify the relevant files. A better workflow selects the code, tests and documentation needed for the current task.

Context length also does not guarantee that a model will understand every file equally well. Evaluate retrieval, file selection, instruction following and change review on the codebase that matters to the team. The coding-agent guide offers a useful comparison framework.

Kimi K2.7 Code API pricing

The official pricing page defines 1M as 1,000,000 tokens and says prices exclude applicable taxes. The cache-hit rate is lower than the cache-miss rate, but eligibility and actual cache behavior depend on the request. Record token usage from the API response rather than assuming every repeated prompt is a cache hit.

Worked Kimi K2.7 Code cost example

Suppose a coding workflow uses 10 million cache-miss input tokens, 20 million cache-hit input tokens and 2 million output tokens in one month. The cache-miss input cost is 10 multiplied by $0.95, or $9.50. The cache-hit input cost is 20 multiplied by $0.19, or $3.80. Output costs 2 multiplied by $4, or $8. The token subtotal is $21.30 before taxes or other charges.

This is a calculation from the published list rates. A real bill may include different cache behavior, retries, account terms, taxes or platform conditions. Measure cost per accepted code change instead of cost per token alone.

Kimi K2.7 Code versus K2.7 Code HighSpeed

Kimi's documentation describes K2.7 Code HighSpeed as the same model with a higher output-speed target. It reports approximately 180 tokens per second and up to 260 tokens per second in short-context scenarios. The documentation also says the HighSpeed resource is limited and availability is being increased gradually.

Speed is not the same as quality or lower total cost. A faster model can improve interactive developer experience, but it may have different resource limits or pricing. Check the current product page before selecting HighSpeed for a production workflow.

Multimodal input for coding work

Kimi K2.7 Code supports text, image and video input according to the pricing documentation. Images can help with interface screenshots, diagrams and visual bugs. Video can provide a recording of a reproduction case. The application should still convert visual observations into testable requirements and keep sensitive media within approved data controls.

Multimodal input can add processing cost or change the token count. Do not estimate a visual task using text-only token math without checking the provider's billing definition.

Thinking mode and long coding tasks

The official model description lists thinking mode and long thinking or deep reasoning support. These features can help with planning, debugging and multi-step changes, but they can also increase latency and output usage. Set an output limit and ask for a clear plan, changed files and test results.

For difficult work, separate planning from execution. Ask the model to identify assumptions first, then let the agent edit only approved files. Review the diff and run the project's tests outside the model response.

Tool calls, JSON Mode and Partial Mode

ToolCalls allow an application to connect Kimi to controlled functions such as file inspection, test execution or issue tracking. JSON Mode can make structured extraction easier when the response must follow a defined schema. Partial Mode can support incremental or constrained response patterns where the product requires them.

Every tool should have a narrow permission boundary. A coding agent should not receive unrestricted production credentials or destructive shell access simply because it can call tools. Log the tool name, arguments, result and approval state for each run. The agentic AI guide covers this control problem. Compare this with current API integration patterns before deployment.

How automatic context caching changes workflow design

Automatic context caching can reduce the price of repeated input when the request qualifies for a cache hit. It is useful for stable instructions, repository maps and repeated documentation. It is not a substitute for prompt discipline. Changing the cached content or request structure can reduce cache reuse.

Track cache-hit and cache-miss tokens separately. A monthly forecast that applies $0.19 to every input token will be wrong if the application does not achieve the expected hit rate. Use measured logs from a representative coding session before setting a budget.

Kimi K2.7 Code for coding agents

Kimi's official quickstart describes the model as dedicated to coding and says it follows instructions more reliably in long contexts. The page also discusses external benchmark evaluations and a reduction in overthinking tendencies. Those statements are provider documentation claims, not an independent guarantee of performance on your repository.

A useful pilot includes issue triage, small bug fixes, feature work, refactoring and test repair. Score the number of accepted changes, review time, failed tests, reverted patches, context used and total cost. Compare against the model already used by the team.

Open model claims and deployment decisions

The previous article described Kimi K2.7 Code with broad parameter and open-weight claims that were not verified in the official pages fetched for this update. They are not repeated here. A deployment decision should use the current API documentation, model identifier, pricing page and any separate release record that can be checked directly.

Hosted API access may reduce infrastructure work, while self-managed access can offer different data and cost controls. Confirm licensing, weights, supported runtimes and operational requirements before treating a model as self-hostable. The compact model deployment guide provides a comparable hosted-versus-local framework.

How to evaluate Kimi K2.7 Code safely

Start with a private test repository or a controlled branch. Give the agent read access first, then allow narrowly scoped edits. Require a plan, a diff and test output. Review every dependency change and do not allow secrets or production credentials in the context.

Measure coding quality and cost together. A lower token bill is not useful if review time rises. A faster response is not useful if the patch fails tests. The right result is a repeatable workflow with acceptable cost, quality, security and latency.

Kimi K2.7 Code 2026 decision guide

Kimi K2.7 Code is worth testing when a team needs a coding-focused model with long context, multimodal input, thinking, tool calls and structured output. Its official list pricing is $0.19 per million cache-hit input tokens, $0.95 per million cache-miss input tokens and $4.00 per million output tokens, excluding applicable taxes.

Use measured repository tasks to decide whether the model fits. Keep the cache rate separate from cache misses, validate all tool permissions and check the current Kimi pricing page before deployment. Model availability, speed tiers, limits and terms can change after this research date.