Skip to content
CheapestInference

Kimi K3.1 Pro API — pricing, access & subscription

Kimi K3.1 Pro is Moonshot’s post-trained successor to Kimi K3, released on October 23, 2026 — it scores 55 on the Artificial Analysis Intelligence Index (v4.3.2), #4 overall and 3 points off the leader (Claude Opus 5.5, 58), one point behind Claude Sonnet 5.5 (56) at the same $2/$10 list price. Its open weights were published on November 4, 2026 under the Kimi K3.1-Pro License (Modified MIT), which permits commercial serving. As of November 6, 2026 CheapestInference serves it in the Flagship Pool, where it replaces Kimi K3 in place next to Qwen3.8 Max, on unlimited, flat-rate time-block subscriptions from $199/mo — no per-token billing, no token caps during your reserved hours. Seats are very limited; live availability is on /pools.

Quick facts

Section titled “Quick facts”
ModelKimi K3.1 Pro
Model IDkimi-k3.1-pro
Context window1M tokens (1,048,576) as served here
Max request size8 MB per request
InputText and image — images as base64 data URLs (OpenAI image_url parts or Anthropic image blocks); public image URLs are not accepted
PoolFlagship Pool (with Qwen3.8 Max); replaced Kimi K3 on November 6, 2026 — kimi-k3 is served by K3.1 Pro until December 7, 2026
Per-token price elsewhere$2.00 in / $10.00 out per 1M (cached input $0.20) — reference only
Endpoints/v1/chat/completions (OpenAI), /anthropic/v1/messages (Anthropic)
ReasoningOn demand — off unless you enable it. reasoning_effort: low, high, max; enabled without an effort = max
Sampling paramsFixed at the model’s calibrated values. Like most reasoning models, Kimi K3.1 Pro is trained and evaluated at a specific sampling configuration — output quality degrades outside it — so every request runs at the calibration behind its benchmark scores. temperature, top_p, presence_penalty and frequency_penalty are ignored if sent
PricingFrom $199/mo — reserve an 8-hour daily time block, up to full 24/7; $169.15/mo with annual billing

Call Kimi K3.1 Pro

Section titled “Call Kimi K3.1 Pro”
from openai import OpenAI


client = OpenAI(
    base_url="https://api.cheapestinference.com/v1",
    api_key="sk-..."  # your subscriber key
)


response = client.chat.completions.create(
    model="kimi-k3.1-pro",
    messages=[{"role": "user", "content": "Refactor this function..."}],
    reasoning_effort="low",  # low | high | max — omit it for a direct answer with no reasoning
    max_tokens=32000,
)
Terminal window
curl https://api.cheapestinference.com/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "kimi-k3.1-pro", "reasoning_effort": "low", "messages": [{"role": "user", "content": "Hello"}]}'

Reasoning is on demand for Kimi K3.1 Pro, as for every model on this API: with nothing set, the model answers directly, with no reasoning tokens. Turn it on per request with "thinking": { "type": "enabled" } or with reasoning_effort — low, high or max. Enabled without an effort, it reasons at max, the setting behind its benchmark scores; send low for interactive or latency-sensitive work (chat, quick edits, short tool calls), and set a generous max_tokens so answers are not truncated mid-reasoning. reasoning_effort: "none" keeps reasoning off. Note that Moonshot’s own API documents K3.1 Pro as always reasoning; here the default is the same as it was for Kimi K3 — off unless you ask — so existing K3 integrations behave the same way.

Images go in as base64 data URLs — OpenAI image_url parts with a data:image/...;base64,... URL, or Anthropic base64 image blocks. Public http(s) image URLs are not accepted by this model (the request fails); download and inline the image instead. Qwen3.8 Max, in the same pool, accepts both forms.

Why unlimited Kimi K3.1 Pro

Section titled “Why unlimited Kimi K3.1 Pro”

At $10 per 1M output tokens elsewhere, and with reasoning on for the agentic work it is best at, Kimi K3.1 Pro is exactly the profile where a flat fee pays off. A Flagship Pool subscription gives you K3.1 Pro through a standard OpenAI-compatible endpoint with no token caps: heavy agentic coding runs, long sessions over its 1M-token context, full workdays, all at one fixed monthly price. It works in Claude Code, Cline, Roo Code, Continue, and any client that accepts a custom base URL.

Because seats are very limited, the pool can sell out — when it does, the pools page shows live availability per block.

Common questions

Section titled “Common questions”

Is there an unlimited Kimi K3.1 Pro API? Yes. The CheapestInference Flagship Pool serves Kimi K3.1 Pro with uncapped usage for one flat monthly fee — from $199/month for a daily 8-hour block ($169.15/month with annual billing), up to full 24/7 with all three blocks. Seats are very limited.

What is the cheapest way to use Kimi K3.1 Pro? For heavy use, an unlimited time-block subscription: from $199/month you reserve a daily 8-hour window and use K3.1 Pro with no token caps. With reasoning on, its output-token bill at the per-token list price ($2 in / $10 out per 1M) grows fast under a coding-agent workload. For light or occasional use, per-token access via Moonshot’s API is the cheaper route.

Is Kimi K3.1 Pro OpenAI-compatible? Yes. Use model id kimi-k3.1-pro against https://api.cheapestinference.com/v1 with the OpenAI SDK, or the /anthropic endpoint with the Anthropic SDK — which also makes it work in Claude Code (set ANTHROPIC_MODEL to kimi-k3.1-pro).

Is there a Kimi K3.1 Pro coding plan? Yes — a Flagship Pool subscription works as a K3.1 Pro coding plan: unlimited Kimi K3.1 Pro through Claude Code, Cline, Roo Code or any OpenAI-compatible coding agent, with no weekly request quotas to manage.

Kimi K3.1 Pro vs Kimi K3 — which one do I get? K3.1 Pro. On November 6, 2026 it replaced Kimi K3 in place in the Flagship Pool — same subscription, same key, nothing to change: requests with model id kimi-k3 are served by Kimi K3.1 Pro until December 7, 2026, after which the old id returns an invalid-model error, so switch to kimi-k3.1-pro. K3.1 Pro scores 55 on the index against K3’s 44, lifts the served context from 256K to 1M tokens, adds image input, and answers directly unless you enable reasoning — exactly as K3 did.

Can I turn reasoning off on Kimi K3.1 Pro? It is off by default: send nothing (or reasoning_effort: "none") and the model answers directly. Enable it per request with thinking: {"type": "enabled"} or reasoning_effort (low for the fastest reasoning, max for the benchmark setting).

Section titled “Related”