MoeRouter Blog
English
Get an API key

Kimi K3.1 Pro review roundup: scores close to Opus 5.5, and split opinions

Moonshot AI's Kimi K3.1 Pro launched on October 24, 2026. Its own scores put it above GPT-6 Astra and close to Claude Opus 5.5 on coding agents. Vendor results, an independent index and Bilibili creators' divided hands-on tests.

kimi-k3.1-pro

In short

  • Kimi K3.1 Pro is the 3-trillion-parameter-class model Moonshot AI released on October 24, 2026, with a 1M-token context. List price is ¥13 in / ¥68 out per 1M tokens, about a third below its predecessor K3.
  • Moonshot reports 65.7% on Terminal-Bench 4.0, above GPT-6 Astra's 58.2% and 0.7 points behind Claude Opus 5.5's 66.4%. Artificial Analysis scores it 55 on its Intelligence Index.
  • Hands-on opinion on Bilibili is split: one reviewer calls it the overall leader among Chinese models, others call it below expectations or a small letdown. The weak spots are scientific computing, some knowledge reasoning and 3D reconstruction.
  • kimi-k3.1-pro was 11th by usage on MoeRouter over the last 7 days. It suits long coding and agent runs; on a tighter budget, try the cheaper Flash tier first.
On this page
  1. What Kimi K3.1 Pro is
  2. What it costs on MoeRouter
  3. Benchmarks
  4. Hands-on: opinion is split
  5. Whether to use it
  6. How to call it
  7. Frequently asked questions

What Kimi K3.1 Pro is#

Kimi K3.1 Pro is the model Moonshot AI released on October 24, 2026: the first of the K3.1 line and, in the company's words, its most capable 3-trillion-parameter-class model so far, with a 1M-token context. A Highspeed variant launched with it: the same intelligence at roughly two to three times the output speed and twice the price. Moonshot said the full model weights would be published by November 4[1][3].

Kimi K3.1 Pro
List price per 1M tokens ¥13 input (¥1.3 on a cache hit), ¥68 output
Against its predecessor K3 was ¥20 input and ¥100 output: input 35% lower, output 32% lower
Context 1M tokens
What Moonshot says it is best at Code, 3D creation, financial research and other long-horizon agent work
Model name on MoeRouter kimi-k3.1-pro
Endpoints on MoeRouter Chat Completions, Messages, Gemini

Prices are as Tencent News and Juejin report them[1][2].

What it costs on MoeRouter#

per 1M tokens
ModelGroupInputOutputCache read
kimi-k3.1-prodefaultCN¥1.3≈ US$0.195CN¥6.8≈ US$1.02CN¥0.13≈ US$0.0195

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

It is in the default group, which shows the reference price of the cheapest route at the moment; the usage log shows what each request was charged.

Benchmarks#

Reported by Moonshot#

Benchmark K3.1 Pro GPT-6 Astra Claude Opus 5.5
Terminal-Bench 4.0 65.7% 58.2% 66.4%
SWE-Marathon v1.1 (long software-engineering tasks) 51.7% 42.5% 46.3%
Agents' Last Exam 43.2 32.8 42.6
BenchCAD (3D reconstruction, with tools) 92.7% 95.9% 96.2%

More of Moonshot's own figures: 61.2% on its internal long-horizon coding benchmark, Kimi Code Bench 3.0, against K3's 55.95%; 64.4 on OfficeQA Pro v2, a benchmark on financial documents, the best result on that test so far; and on an internal stress test of factual accuracy, the error rate per factual statement fell from K3's 23.4% to 8.3%[1][2].

Independent#

Artificial Analysis scores K3.1 Pro 55 on its Intelligence Index, against 44 for K3. It says that once the weights are open this will be the highest-scoring open-weights model, narrowing the gap between open weights and closed models from 12 points to 3[2].

Where it is weaker#

The Juejin author is plain about it: on scientific-computing terminal tasks, some knowledge reasoning and 3D reconstruction, GPT-6 Astra and Claude Opus 5.5 still lead. Moonshot also removed content about network attacks from its training data on purpose; its internal cybersecurity benchmark fell from 90.5 to 81.0 as a result, with general coding ability almost unaffected[2].

Hands-on: opinion is split#

Unlike the scores, the hands-on verdicts on Bilibili fall on two sides.

小宇Boi, a week after launch and with 200,000 views, is positive: K3.1 Pro leads Chinese models on overall performance and did very well in his tests[4].

实验式 titled his video "below expectations". He holds the vendor's claims, 3T parameters, a 1M context and scores that beat GPT-6 Astra on several measures, against how it actually behaves[5].

等二月花开 ran the same task on Opus 5.5 and writes that K3.1 was a small letdown at first, though next to what Opus 5.5 made the gap did not seem that large[6].

盒子姬 tested Qwen 4 and Kimi K3.1 Pro together in Godot and rated both "NPC", tier-list slang for unremarkable, at about ¥19 for Qwen and ¥24 for Kimi[7]. The same creator had rated Claude Fable 5.2 top tier; see the Fable 5.2 review.

Put together: the scores show that it has reached the first rank on standardised coding-agent tasks. These creators tested projects of their own choosing, Godot game development for one, which measure something else. And Moonshot made 3D creation and game engines the showpiece of this launch, which raised expectations and is part of the letdown.

Whether to use it#

Good for

  • Long coding and agent runs
  • Financial documents: analysis, cross-checking and multi-step numerical reasoning
  • Flagship-class coding without a Claude flagship's price

Less suited to

  • Scientific-computing terminal tasks
  • Projects that expect a great deal of its 3D and game visuals
  • High-volume simple work: the Flash tier is better value
If you Then
Want a cheaper alternative to Claude Opus 5.5 It is worth a round on your own tasks: the vendor's score is 0.7 points behind
Mostly build 3D and game prototypes Try it on a small scale first; hands-on results are uneven
Are on a tight budget Start with DeepSeek V4.1 Flash or GLM-5.3-Flash and move up when they are not enough

How to call it#

On MoeRouter it answers on the Chat Completions, Messages and Gemini endpoints. With the OpenAI SDK:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://moerouter.com/v1",
    api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
    model="kimi-k3.1-pro",
    messages=[{"role": "user", "content": "Check the revenue figures in these two annual reports against each other and list what does not match"}],
)
print(reply.choices[0].message.content)

In Claude Code:

export ANTHROPIC_BASE_URL="https://moerouter.com"
export ANTHROPIC_AUTH_TOKEN="$MOEROUTER_API_KEY"
unset ANTHROPIC_API_KEY
claude --model kimi-k3.1-pro

It is not offered on the Responses endpoint, so it cannot be used in Codex as it is. The docs are in Chinese for now.

Frequently asked questions#

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

Moonshot's list price per 1M tokens is ¥13 input (¥1.3 on a cache hit) and ¥68 output, and twice that for the Highspeed variant. MoeRouter's reference prices are in the table above.

Is Kimi K3.1 Pro better than Claude Opus 5.5?#

On Moonshot's Terminal-Bench 4.0 figures it is 0.7 points behind Opus 5.5, and ahead on long software-engineering tasks and Agents' Last Exam. Opus 5.5 is stronger on scientific computing, some knowledge reasoning and 3D reconstruction. A side-by-side test on Bilibili called it a small letdown with a gap that was not large.

Is Kimi K3.1 Pro open source?#

At launch Moonshot said the full model weights would be published by November 4, 2026. For whether they are out and under which license, check Moonshot's own pages.

Can I use Kimi K3.1 Pro in Claude Code?#

Yes. On MoeRouter it supports the Anthropic Messages endpoint; start with --model kimi-k3.1-pro. It does not support the Responses endpoint, so it cannot be used in Codex.

Why do the hands-on verdicts differ so much from the scores?#

The scores measure standardised coding-agent tasks, where it is close to Opus 5.5. The Bilibili creators tested projects of their own choosing, Godot game development for one, which measure something else; and Moonshot made 3D creation and game engines the showpiece, so expectations were high.

What is the difference between Kimi K3.1 Pro and K3?#

K3.1 Pro succeeds K3: Moonshot's long-horizon coding benchmark rose from 55.95% to 61.2%, the error rate per factual statement fell from 23.4% to 8.3%, and the price fell by about a third. Its Artificial Analysis Intelligence Index rose from 44 to 55.

Sources

The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.

  1. Kimi K3.1 Pro发布,月之暗面最强3万亿级模型 (in Chinese) Tencent News · 三易生活 · published 2026-10-25
  2. 国产Kimi K3.1 Pro实测解读:3万亿参数逼近Claude,价格只要五分之一 (in Chinese) Juejin · 咬代码的兽 · published 2026-10-25
  3. Kimi K3.1 Pro 来啦:智能前沿,更进一步 (video, in Chinese) Bilibili · Kimi智能助手 (official) · published 2026-10-24
  4. 【实测】Kimi K3.1pro发布一周,真实体验究竟怎样? (video, in Chinese) Bilibili · 小宇Boi · published 2026-10-31
  5. 不及预期, Kimi K3.1 Pro 实测 (video, in Chinese) Bilibili · 实验式 · published 2026-10-25
  6. Kimi K3.1 Pro 首测 | 狂写1.5小时憋坨大的? (video, in Chinese) Bilibili · 等二月花开 · published 2026-10-24
  7. 【盒子姬】Qwen 4 和 Kimi 3.1Pro - Godot 能力评测 (video, in Chinese) Bilibili · 盒子姬 · published 2026-11-14

Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.