Kimi K3.1 Pro: price, benchmarks and whether to switch

Kimi K3.1 Pro is Moonshot AI's newest model, released on 23 October 2026. It is Kimi K3 trained much further with reinforcement learning, keeps the 1-million-token context windowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out., and costs $2 per million input tokens and $10 per million output tokens on the Kimi API. That is a third less than Kimi K3, and the same base price as Claude Sonnet 5.5 and GPT-6.1 Sol. Moonshot says the full weights will be public by 4 November.

The model is better at what K3 already did, especially long coding tasks, and it is cheaper. Moonshot's own benchmarks put it just behind Claude Opus 5.5 and ahead of GPT-6 Astra on agent coding. No one outside Moonshot had scored it at launch, so treat those numbers as a claim to check on your own work.

Kimi K3.1 Pro at a glance

The dates that matter

What changed from Kimi K3

K3.1 Pro is not a new base model. Moonshot describes it as a post-trained version of K3, with much more reinforcement learning compute spread over hundreds of task types: coding, 3D modelling, financial research, office documents. The context window stays at 1,048,576 tokens.

What a developer notices:

• The price went down. $2 and $10 per million tokens against K3's $3 and $15. Cached input drops from $0.30 to $0.20.
• There is a fast version. kimi-k3.1-pro-highspeed answers 2 to 3 times faster, at twice the standard price.
• It always reasons. There is no switch to turn thinking off. You set reasoning_effort to low, high or max, and the default is max. The reasoning tokens are billed, so the default is also the most expensive setting.
• Weights are coming. Moonshot calls it open-weight and says the weights will be out by 4 November. Until then it runs only on Moonshot's own apps and API.

It is in the Kimi app, the Kimi Work desktop app (version 3.1.0 or later), Kimi Code (pick it with /model) and the API. On Kimi Enterprise an administrator has to turn it on, because access is off by default. On the API it is a "flagship" model, unlocked once the account has topped up at least $1.

What it costs next to the models you use now

Per-token prices are easy to compare. Here they are with one sum for a coding-agent step: 200,000 tokens of context, of which 160,000 are already in the prompt cachePrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker., plus 20,000 tokens of output, reasoning included.

Prices are per million tokens, from each maker's own page. The step sum is 40,000 new input tokens, 160,000 cached and 20,000 output.

Three things stand out.

First, K3.1 Pro has the same list price as Sonnet 5.5 and GPT-6.1 Sol but double their cached input price, so in a cache-heavy agent it comes out a little dearer, not cheaper. With no caching at all, the 3 cost the same.

Second, Highspeed costs exactly what Claude Opus 5.5 costs per token. Paying for speed puts it in the price band of a stronger model.

Third, anyone still on K2.6 or K2.7 Code pays about twice as much per step after the move, because the K2 models are being retired (more on that below).

Move the sliders to your own agent. The more of the context comes from the cache, the further K3.1 Pro falls behind Sonnet 5.5 and GPT-6.1 Sol; push the context past 262K tokens and K2.6 drops out, because its window is a quarter of K3.1 Pro's.

Try itWhat one agent step costs on Kimi K3.1 Pro and its rivals

Context sent each step

200K

Read from the cache

80%

Output, reasoning included

20K

Steps a month

10,000

Cost of one step, cheapest first

• DeepSeek V4 Pro$0.139$1,390 a month

• Kimi K2.6$0.1436$1,436 a month

• Claude Sonnet 5.5$0.296$2,960 a month

• GPT-6.1 Sol$0.296$2,960 a month

• Kimi K3.1 Pro$0.312$3,120 a month

• Kimi K3$0.468$4,680 a month

• Kimi K3.1 Pro Highspeed$0.592$5,920 a month

• Claude Opus 5.5$0.592$5,920 a month

The sums leave out one thing: models do not use the same number of tokens for the same job. Tokenizers differ, and a model that reasons longer bills more output. The only honest comparison is your own task run on each model and the bills compared. The cost calculator does the sums for your own token counts, and the Kimi K3.1 Pro price page lists every price, with what common jobs cost on it.

For the caching details: on the Kimi API caching is on by default with a 5-minute lifetime. Writing a prefix into the cache costs $2 per million tokens for 5 minutes or $4 for 1 hour, and every hit starts the clock again. As on every provider, the cache only matches the exact start of a prompt, so a timestamp near the top makes it useless. Our guide to cutting LLM API costs covers how to lay out a prompt so the cache hits.

What Moonshot's benchmarks show, and what they don't

Moonshot says K3.1 Pro scores 65.7% on Terminal-Bench 4.0, 0.7 points behind Claude Opus 5.5 and 7.5 ahead of GPT-6 Astra. It published scores across coding, agent, knowledge and image benchmarks; the chart has the 3 whose figures are written in its post, with the price of the same agent step beside each model.

Try itKimi K3.1 Pro against Claude and GPT on Moonshot's benchmarks

Benchmark

Terminal-Bench 4.0BenchCADKimi Code Bench 3.0

Terminal-Bench 4.0

Agent coding tasks in a terminal, checked by each task's own tests.

• Claude Opus 5.5
  66.4%$0.592 a step
• Kimi K3.1 Pro
  65.7%$0.312 a step
• GPT-6 Astra
  58.2%$1.56 a step

The coding result is the one that matters for most readers: on agent work in a terminal, Moonshot puts K3.1 Pro level with Opus 5.5 at half the price per step. On BenchCAD, rebuilding 3D objects as CAD code, it trails both Opus and Astra by 3 points or more. Kimi Code Bench is Moonshot's own test, so it shows the gain over K3 (61.2% against 55.95%) rather than a ranking.

Moonshot also reports fewer made-up facts. On its own hard factual test the share of false claims fell from 23.4% for K3 to 8.3%. It says that test is built to catch errors and does not give the error rate in everyday use.

How those numbers were made matters as much as the numbers. Moonshot ran its own model at maximum reasoning effort, mostly inside the Claude Code harness. Most rival scores come from their makers' reports or public leaderboards, and 2 of the benchmarks are Moonshot's own. That makes this a strong showing, not an independent ranking.

We show an independent score for every model on /prices: the Epoch Capabilities Index from Epoch AI. Epoch had not scored K3.1 Pro at launch, so its price page shows it as untested for now. Kimi K3 scores 157.42 on that index, against 166.82 for Claude Opus 5.5. K3.1 Pro will very likely score higher than K3, and how much higher is what will show whether it really sits next to Opus.

If you are on Kimi K2

This is the part with a date on it. From 31 October 2026, accounts registered after that day cannot use the K2 models; existing accounts keep them. Moonshot plans to retire all K2 models by the end of December 2026, and its price page points K2 users to K3.1 Pro.

The API call stays OpenAI-compatible, so most code keeps working. Two things change:

1. Remove the K2 thinking setting and set reasoning_effort instead. K3.1 Pro always reasons, so the choice is how hard.
2. Send back the whole assistant message in multi-turn chats and tool loops, including reasoning_content and tool_calls. Code that keeps only content between turns breaks the model's reasoning chain.

import os
from openai import OpenAI

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

messages = [{"role": "user", "content": "Find why tests/test_orders.py fails and suggest a fix."}]
reply = client.chat.completions.create(
    model="kimi-k3.1-pro",        # was kimi-k2.6 or kimi-k2.7-code
    reasoning_effort="high",      # replaces K2's thinking setting; the default is max
    messages=messages,
)
messages.append(reply.choices[0].message)   # the whole message, not only .content

Then look at the bill. The same traffic costs about twice as much per token, and max effort writes the most reasoning tokens. Start agent loops on high or low, and raise it only for the steps that need it.

Running it yourself on 4 November

Open weights mean you can run K3.1 Pro on your own GPUs or pick a host other than Moonshot. Moonshot calls it a "3T-class" model, which means trillions of parameters: it is a multi-GPU server job, not a laptop one. The licence and the exact files come with the weights, so check both before you plan on it. Once the config files are out, we will add it to the GPU memory calculator; until then, how much VRAM a model needs explains the sum.

Moonshot also says it removed cyber-attack material from the training data of the version it will publish. On its internal exploit test the published version solves 81.0% of tasks, against 90.5% for the same training run without that removal. It says the model still trails the leading closed models on this kind of work.

For data teams the API terms matter more than the weights: Moonshot offers Zero Data Retention to API customers, and Kimi Business keeps organisation accounts apart from personal ones.

Should you switch?

• From Kimi K3: yes. The model is newer and better on Moonshot's numbers, and a third cheaper. Change the model id and review your reasoning_effort.
• From K2.6 or K2.7 Code: you will have to. Plan the change before the end of December, and budget about twice the token cost.
• From Claude Sonnet 5.5 or GPT-6.1 Sol: only after a test. The list price is the same and the cached price is higher, so the case rests on quality on your own tasks. Run a few hundred of them on both and compare. The Sonnet 5.5 vs K3.1 Pro page puts the 2 side by side.
• From Claude Opus 5.5: maybe, for cost. Half the token price for a model that Moonshot's figures place just behind it in agent coding. Wait for an independent score, then try it on the steps where Opus is the expensive part of your bill.
• If you need it on your own hardware: wait for 4 November, read the licence, and size the servers first.

Not sure which model fits a job? Which model should I use starts from the job and the budget.