I investigated which Kimi API model to choose for enterprise-oriented code review, contract clause extraction, and requirements validation. On November 10, 2026, I sent the same task to the eight model IDs listed in the API directory that day. Only three models—kimi-k2.7-code-highspeed, kimi-k2.7-code, and kimi-k2.5—returned complete JSON. The high-speed version finished in 8.48 seconds, while the current flagship, kimi-k3.1-pro, hit the 180-second timeout.

This article is for developers who want to handle structured data via an API. Because this was a text-only, one-off request, it is not an evaluation of the 1-million-token context, image and video input, or coding agents that manipulate repositories.

Kimi API Conclusion: Try These 3 Models First

Task	First Model to Try	Result This Time	What to Check Next
Receive structured data quickly	kimi-k2.7-code-highspeed	Complete JSON in 8.48 seconds	Availability period for the route
Code review and requirements verification	kimi-k2.7-code	Complete JSON in 36.53 seconds. Eight items remained unverified.	Transition timing and current account permissions
Establish comparison criteria against general-purpose models	kimi-k2.5	Complete JSON in 32.26 seconds	No longer available on the official platform
Longer code and document reviews	kimi-k2.8-preview	The code and contract were returned, but the final JSON was cut off partway through	Adjusting task division and output limits
Output long reviews in a single batch	kimi-k3, kimi-k3.1-pro-highspeed, kimi-k2.6	Reached the output limit, with no complete JSON	Shorten the output of one stage
Evaluate the current flagship	kimi-k3.1-pro	Timed out after 180 seconds	Retest with a shorter request and timeout

Based solely on this result, I would choose kimi-k2.7-code-highspeed for the initial connectivity check. However, K2.5 has been discontinued on the official platform, and K2.6 and K2.7 Code are no longer available to newly registered users. The presence of IDs in the API directory and the official availability period must be checked separately.

Kimi API testing conditions and evaluation method

The request conditions are the same for all models.

• Non-streaming Chat Completions request
• max_tokens=5000
• temperature is unset
• Once per model
• 180-second timeout
• Use the same enterprise development task

The task consisted of three parts: analyzing a JavaScript bug whose discount rate changes according to array order, extracting fields from a synthesized procurement contract, and validating overnight batch requirements involving duplicate orders, re-execution, and audit logs.

The evaluation is calculated as follows: code analysis 30%, document extraction 25%, requirements review 25%, and format/invocation reliability 20%. A score of 0 means that there was no evaluable output under the conditions of this test; it does not mean that the model’s general capabilities were scored as 0.

Complete results for all 8 Kimi API models

Model IDs listed in the API directory that day	Score	Evaluation	Actual response
kimi-k2.7-code	4.9/5	★★★★★	Complete JSON. Covered the code, contract, and requirements
kimi-k3	0.0/5	★☆☆☆☆	Returned HTTP 200, but reached the output limit, with no complete deliverable
kimi-k2.8-preview	3.6/5	★★★☆☆	The code and contract were returned, but the JSON was cut off midway
kimi-k2.5	4.7/5	★★★★★	Complete JSON. A result usable for all three tasks
kimi-k3.1-pro	0.0/5	★☆☆☆☆	Timed out. No task output available for evaluation
kimi-k3.1-pro-highspeed	0.0/5	★☆☆☆☆	Reached the output limit; no complete JSON
kimi-k2.6	0.0/5	★☆☆☆☆	Reached the output limit; no complete JSON
kimi-k2.7-code-highspeed	4.9/5	★★★★★	Complete JSON. The fastest response time in this test

Seven models returned HTTP 200, but complete JSON was obtained from only three models. I believe this is the difference between a mere connectivity check and practical usability. Even when the status is successful, the content may not be processable by a downstream parser.

Kimi K2.7 Code HighSpeed: Complete JSON in 8.48 seconds

The high-speed version scored the same 4.9 as kimi-k2.7-code, with a response time of 8.48 seconds in this test. It was not made faster by omitting output; it includes all three sections: code review, contract extraction, and requirements validation.

For the code issue, it identified the implementation that processes the array sequentially and overwrites the discount rate as the cause, and listed four types of regression cases: out-of-order tiers, amounts exactly at the boundary value, no applicable tier, and invalid rules.

For the contract, it correctly distinguished the total amount of 186,500 yuan, the advance payment of 55,950 yuan, and the balance of 130,550 yuan, and correctly extracted the contract date, both parties, and the three-year warranty.

For the requirements, it identified a conflict between “keep the last record for duplicate IDs” and “keep the first ingested record for auditing,” and also detected the processing-time constraint of approximately one hour, from 23:00 to 00:00 the following day. After confirming these three points, I judged it sufficient to use as the first candidate for validation.

Kimi K2.7 Code and Kimi K2.5: Differences in Intended Use

kimi-k2.7-code took 36.53 seconds, but the depth of its requirements review differed from that of the high-speed version. In addition to duplicate records and processing time, it also noted the idempotency required to rerun failed records and prevent double payments, and listed eight items requiring confirmation. This is the model I would use to recheck code and requirements.

kimi-k2.5 completed in 32.26 seconds, and its code and contract results were accurate. For the requirements, it explicitly identified one conflict and three items requiring confirmation, and the output was 4,930 tokens. Although the output contained more characters than the high-speed version, its coverage of the requirements was narrower, so I treat it as a baseline for comparison with general-purpose models.

However, this differentiation is based on the response content from this test.K2.5 has been discontinued, and K2.7 Code is in a transition period, so whether it can be used long-term is a separate verification item.

Kimi K2.8 Preview and the Kimi K3 Series: Unusable Responses Despite HTTP 200

For kimi-k2.8-preview, the code and contract contents could be confirmed, but the JSON was cut off midway through the requirements review. Since the portion already returned is useful, it is a candidate for dividing the task and revalidating it.

kimi-k3, kimi-k3.1-pro-highspeed, and kimi-k2.6 reached the 5,000-token output limit, and their content was empty. kimi-k3.1-pro timed out after 180 seconds.

In light of these results, I check the following items before accepting a response.

1. Is the content non-empty?
2. What is the finish_reason?
3. Can it be parsed as JSON?
4. Are required fields and business values, such as amounts, correct?

If HTTP 200 alone is used as the completion condition, empty responses or objects cut off midway may be passed downstream.

The Kimi Model Lifecycle Is Also a Selection Criterion

Date January 2026 April 2026 June 2026 July 2026 September 11, 2026 October 23, 2026 As of November 10, 2026

Model/changes	Current implications
K2.5 launched on the official API	Now discontinued
K2.6 launched on the official API	Availability to newly registered users discontinued
K2.7 Code and the high-speed version launched	Full discontinuation planned for the end of December 2026
K3 launched on the official API	Previous-generation flagship; 1-million-token context
K2.8 Preview made fully available in Kimi Code	The client-side ID is kimi-for-coding
K3.1 Pro and Highspeed announced	Current flagship
K3.1 Pro weights and license released	Already released, not scheduled for release

I record separately whether it could be invoked this time and whether it can be used in the future. This success does not extend the official availability period.

Kimi K3.1 Pro Specifications and Public Benchmarks

K3.1 Pro is a model with enhanced post-training over K3. Moonshot AI explains that it improved the reinforcement learning framework, long-context reasoning, sandboxing, and scheduling for coding, 3D production, financial research, and long-running agent tasks.

Item	Official specifications
Architecture	Mixture of Experts
Total parameters	2.8 trillion
Active parameters per token	104 billion
Layers	93
Experts	896, with 16 selected per token
Context	1,048,576 tokens
Model card modalities	Text and images. The API guide also includes a description of video input.
Reasoning	Always enabled. low, high, or max can be specified
Quantization	MXFP4 weights, MXFP8 activations
Official API IDs	kimi-k3.1-pro, kimi-k3.1-pro-highspeed

In the public benchmarks, Terminal-Bench 4.0 is 65.7%, SWE-Marathon v1.1 is 51.7%, BenchCAD with tools is 92.7%, OfficeQA Pro v2 is 64.4%, and DeepSWE v1.1 is 75.5%. These are figures from the official model card and are separate from my API verification scores.

The Decoder’s October 23, 2026 article reports that the Artificial Analysis Intelligence Index was 55, 11 points higher than K3, and ranked fourth at the time. This was an external observation at the time of publication, not a permanent ranking.

Pricing for the Kimi API and Kimi Code

These are the fees listed on Kimi’s official open platform, checked on November 10, 2026.

Official API Model	Input	Output	Cache Hit
kimi-k3.1-pro	13 yuan	68 yuan	1.30 yuan
kimi-k3.1-pro-highspeed	26 yuan	136 yuan	1.30 yuan
kimi-k3	20 yuan	100 yuan	2.00 yuan
kimi-k2.7-code	6.50 yuan	27 yuan	1.30 yuan
kimi-k2.7-code-highspeed	13 yuan	54 yuan	2.60 yuan
kimi-k2.6	6.50 yuan	27 yuan	1.10 yuan

The K3 series has separate cache-write TTLs of 5 minutes and 1 hour, and the Batch APIs for K3.1 Pro and K3 are priced at 60% of the standard rates. The input and output unit prices for K3.1 Pro are about one-third lower than those for K3, but total usage varies depending on inference output, caching, retries, and tool calls.

Kimi Code is a separate membership program. The new plans are Plus at 99 yuan per month, Pro at 199 yuan per month, Max at 699 yuan per month, and Ultra at 1,399 yuan per month; the monthly equivalents for annual billing are 79 yuan, 159 yuan, 559 yuan, and 1,119 yuan, respectively. Go does not include a Kimi Code usage allowance, but Kimi Code can be used from the Plus plan onward. Highspeed and a maximum context window of 1M are available on Pro and above. Membership allowances and Open Platform API charges cannot be combined.

Open Weights and Access Options for Kimi K3.1 Pro

The official Hugging Face moonshotai/Kimi-K3.1-Pro makes the weights and Kimi K3.1 Pro License available. I reviewed the model card and license but did not download or locally deploy it. The official guidance covers vLLM, SGLang, TokenSpeed, and other tools, but running a 2.8-trillion-parameter model requires calculations for storage capacity, VRAM, and inference costs.

Ways to access it vary by use case.

• **Kimi Web and app:** General conversations and document processing.
• **Kimi Work:** K3.1 Pro is available in version 3.1.0 and later.
• **Kimi Code:** Select a model on the desktop, CLI, or VS Code. Usage conditions vary depending on membership tier.
• Kimi Open Platform:Uses kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Reasoning is always enabled.
• **Aggregation API:** This time, Crazyrouter’s OpenAI-compatible endpoint was used, with the Base URL set to https://api.crazyrouter.com/v1. Model IDs can be checked in the Crazyrouter model directory. To compare the same prompt, test it first in the comparison environment, then verify the results in your own application as well.

FAQ on Choosing a Kimi Model

Does a score of 0 for Kimi K3.1 Pro mean that the flagship cannot be used?

No. It means that there was nothing that could be evaluated within 180 seconds this time. The official specifications, public benchmarks, and results of a single API call are separate pieces of evidence.

Can reasoning be disabled for Kimi K3.1 Pro?

No. The official API documentation states that it is always enabled. You can change the reasoning effort, but you cannot disable reasoning itself.

Kimi K2.5 received a high score, but can it be used for a new project?

I would not decide based solely on this time’s score. It returned complete JSON this time, but it is no longer available on the official platform. You need to check separately how long it will remain available through the Aggregation API.

Can I directly enter kimi-k2.7-code into Kimi Code?

Do not copy it as-is. The IDs displayed in Kimi Code are k3.1-pro, k3.1-pro-highspeed, k3, k3-256k, kimi-for-coding, and kimi-for-coding-highspeed. Check the client ID and the Aggregation API route ID separately.

If you had to choose one based solely on this time’s results?

I would choose kimi-k2.7-code-highspeed because it returned complete JSON in 8.48 seconds. For long-term projects, check the route’s availability period and also re-evaluate K3.1 Pro using short tasks.

References for Kimi Models

• Kimi K3.1 Pro Official Technical Blog: Development background, public benchmarks, and safety considerations.
• Kimi Official Model List and Changelog: Model status and version history.
• Official Kimi API Pricing: Pricing for input, output, caching, and Batch.
• Kimi K3.1 Pro's Hugging Face repository: Weights, model card, and license.
• Kimi Code model settingsandmembership rules: Client model IDs and membership requirements.
• The Decoder's release-day report: Observations from Artificial Analysis as of publication.