Kimi K3.1 Pro API Guide (2026): $2/$10 Pricing, Reasoning Effort, and K2 Migration

On October 23, 2026, Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series, and put it on the Kimi API the same day. The headline numbers: a 1,048,576-token context window, a price of $2.00 per million input tokens and $10.00 per million output tokens, and a promise that the open weights arrive by November 4. Moonshot's own Terminal-Bench 4.0 score is 65.7%, 0.7 points behind Claude Opus 5.5.

That's a lot of launch-day claims. This guide sticks to what you can check on the pricing page and in the docs.

What Moonshot actually shipped

Two API models went live. kimi-k3.1-pro is the main one. kimi-k3.1-pro-highspeed is the same model at roughly 2-3x the output speed, for exactly double the price. Both get the 1,048,576-token window and native visual understanding.

Moonshot describes K3.1 Pro as a post-trained version of K3 with much more reinforcement-learning compute, aimed at long-horizon coding, 3D and CAD generation, and financial research. The benchmark claims worth knowing: 65.7% on Terminal-Bench 4.0, 61.2% on Moonshot's internal Kimi Code Bench 3.0 Long-horizon (K3 scored 55.95%), and 92.7% on BenchCAD with tools. All of them come from Moonshot's launch post, so treat them as a reason to run your own eval.

Kimi API pricing: what each model costs

This is the rendered table from the Kimi API Platform pricing page on October 24, 2026. All prices are USD per 1M tokens, before tax. Source: Kimi API Platform pricing.

Two details matter. First, the table lists one rate per model across the whole window. There's no separate long-context tier like the one GPT-6.1 Sol applies above 272K input tokens. Second, caching is automatic with a 5-minute default TTL, and cache writes are priced as their own line: $2.00 per 1M tokens for a 5-minute write and $4.00 for a 1-hour write on K3.1 Pro. Cache hits cost $0.20, a tenth of the miss price.

Calling kimi-k3.1-pro

The Kimi API accepts Chat Completions, Responses, and Messages formats, so the OpenAI SDK works with a base URL swap. One gotcha: flagship models unlock after a successful top-up of at least $1.

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kimi-k3.1-pro",
    "reasoning_effort": "high",
    "messages": [
      {"role": "user", "content": "Find the N+1 query in this handler and fix it."}
    ]
  }'

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["MOONSHOT_API_KEY"],
    base_url="https://api.moonshot.ai/v1",
)

resp = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="high",
    messages=[{"role": "user", "content": "Find the N+1 query in this handler and fix it."}],
)
print(resp.choices[0].message.content)
print(resp.usage)  # log this: it's how you'll price thinking tokens

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.MOONSHOT_API_KEY,
  baseURL: "https://api.moonshot.ai/v1",
});

const resp = await client.chat.completions.create({
  model: "kimi-k3.1-pro",
  reasoning_effort: "high",
  messages: [{ role: "user", content: "Find the N+1 query in this handler and fix it." }],
});
console.log(resp.choices[0].message.content, resp.usage);

Thinking is always on, so reasoning_effort is your cost dial

K3.1 Pro has no non-thinking mode. Moonshot's docs say it plainly: you can't turn thinking off, and if reasoning takes too long, set reasoning_effort to low. The accepted values are low, high, and max, and max is the default.

That default is the trap. Migrate a K2.6 or K2.7 Code workload by changing only the model name and every request now runs at max effort. The docs have a section on reasoning-token billing, so assume thinking tokens land on the output side at $10.00 per million, and read them from the usage field instead of guessing.

Here's why that matters. Take a daily mix of 10M uncached input, 40M cached input, and 5M output tokens. At list prices, Kimi K3.1 Pro costs $78.00 and GPT-6.1 Sol costs $74.00. Output alone is $50.00 of each. The $4.00 gap is noise; the effort setting moves the bill more. DeepSeek-V4.1-Flash-1017 at off-peak rates comes to $4.62 on the same mix, about 17x cheaper, though it's a different class of model.

My default: high for agent loops, low for routing and summarization steps, max only for the genuinely hard planning call. Measure before you believe me.

Picking a cache TTL

A 1-hour write costs $2.00 more per 1M tokens than a 5-minute write, and every hit saves $1.80 versus a miss. If your agent pauses longer than five minutes between turns (human review, CI waits), two reuses inside the hour save $3.60 against that $2.00 premium. If turns arrive faster than that, leave the default alone.

Kimi K3.1 Pro vs Claude Sonnet 5.5, GPT-6.1 Sol, and DeepSeek-V4.1-Flash

Prices for the non-Kimi rows come from each provider's pricing page as checked earlier this month; confirm them before you commit budget.

Kimi K3.1 Pro matches Claude Sonnet 5.5 and GPT-6.1 Sol dollar for dollar on input and output, then charges twice as much for cached reads. So the case for it isn't price. It's the open-weights commitment, a flat 1M window with no surcharge tier, and a coding-agent score that, if it holds on your repos, sits near the top of the market. If it doesn't hold, you've lost an afternoon.

The K2 deadline nobody put in the headline

Starting October 31, 2026, K2 series models (kimi-k2.7-code, kimi-k2.7-code-highspeed, and kimi-k2.6) are no longer available to newly registered users. Existing users aren't affected yet, but Moonshot plans to retire the whole series by the end of December 2026. The older kimi-k2.5 and moonshot-v1 models were already discontinued on August 31, 2026.

Moving from K2.7 Code ($0.95 input, $4.00 output) to K3.1 Pro ($2.00 and $10.00) is a 2.1x increase on input and 2.5x on output, plus always-on thinking. Budget for that now, not the week your endpoint disappears.

A sane rollout

Run K3.1 Pro as a challenger, not a replacement. Replay a week of real agent traces against it at reasoning_effort: "high", log tokens and pass rates, and compare cost per solved task rather than cost per token. If you already send Claude and GPT traffic through an OpenAI-compatible gateway like KissAPI, keep that path as your control group and call Kimi directly as the challenger, so the comparison stays honest. Price your own mix with the API cost calculator, and check prompt sizes with the token counter before you pick a cache TTL.

If open weights land on November 4 as promised, self-hosting becomes a real option for teams with data-residency constraints.

FAQ

How much does the Kimi K3.1 Pro API cost?

Kimi K3.1 Pro costs $2.00 per 1M input tokens, $10.00 per 1M output tokens, and $0.20 per 1M cached input tokens, with a 1,048,576-token context window. The kimi-k3.1-pro-highspeed variant costs $4.00 per 1M input tokens and $20.00 per 1M output tokens for roughly 2-3x the output speed.

Can you turn off thinking in Kimi K3.1 Pro?

No. Kimi K3.1 Pro always runs with thinking enabled. You can change the reasoning_effort request field to low, high, or max (max is the default) to trade reasoning depth against latency and token usage.

When are the Kimi K2 models being retired?

Starting October 31, 2026, Kimi K2 series models (kimi-k2.7-code, kimi-k2.7-code-highspeed, and kimi-k2.6) are no longer available to newly registered users. Existing users are not affected yet, and Moonshot plans to retire the whole K2 series by the end of December 2026.