Kimi K3.1 Pro Review: Benchmarks, Pricing and What's New

Kimi K3.1 Pro is Moonshot AI's new flagship, released October 23, 2026. It is Kimi K3 with far more post-training, the same 1M-token context, and a lower price: $2 per million input tokens and $10 per million output. On Moonshot's 14-test chart it beats Claude Opus 5.5 on 9 tests and GPT-6 Astra on 10.

An independent check mostly agrees: Artificial Analysis ranks it 4th of 215 models, 3 points behind Opus 5.5, and its own Terminal-Bench 4.0 run gave K3.1 Pro 67.2%, more than Moonshot claimed.

The catches: Artificial Analysis finds it slower than average and very verbose, and the open weights are due by November 4. We haven't run K3.1 Pro ourselves yet, so the numbers below are Moonshot's and Artificial Analysis's.

Short answer

• A third cheaper than K3, same API. $2 in and $10 out per million tokens against K3's $3 and $15; you switch by changing the model name to kimi-k3.1-pro.
• Close to Claude Opus 5.5 on Moonshot's chart. 65.7% on Terminal-Bench 4.0 against 66.4% for Opus 5.5 and 58.2% for GPT-6 Astra. Its weak spot is science: 38.1% on Terminal-Bench Science against 68.1% for GPT-6 Astra, a gap Artificial Analysis's own run confirms.
• Independent tests back the coding score. Artificial Analysis's own Terminal-Bench 4.0 run gave it 67.2%, ahead of Opus 5.5's 59.6%, and its index ranks it 4th of 215 models. The catch: it runs at 73 tokens a second and writes a lot, so a task costs $2.11, closer to GPT-6 Astra's $3.26 than its token price suggests.
• Our take: on K3, move to K3.1 Pro now. On Opus 5.5 or GPT-6 Astra, run a few of your own tasks first: the savings are real, but most of its scores still come from Moonshot alone.

What is Kimi K3.1 Pro?

It's Kimi K3 trained much further, at a lower price. Moonshot calls it "our most capable 3T-class open-weight model to date" and the first model of a new K3.1 series.

Two model names: kimi-k3.1-pro and kimi-k3.1-pro-highspeed, which Moonshot says is 2 to 3 times faster for twice the price.

Today it runs in the Kimi app and kimi.ai, Kimi Work (3.1.0 or later), Kimi Code (pick it with /model) and Moonshot's API. For Kimi Enterprise it stays off until an admin turns it on. On OpenRouter, Moonshot is the only provider so far.

How does it score against Claude Opus 5.5 and GPT-6 Astra?

On Moonshot's own chart it trades wins with both: it beats Claude Opus 5.5 on 9 of 14 tests and GPT-6 Astra on 10.

Scores are percentages except GDPval-AA, an Elo rating; from Moonshot's launch post, October 23, 2026.

Against Opus 5.5, its biggest leads are on SWE-Marathon (51.7 against 46.3) and GDP.pdf (34.6 against 30.6). Its clearest loss is Terminal-Bench Science: 38.1 against 68.1 for GPT-6 Astra and 63.3 for Opus 5.5.

It also beats DeepSeek V4.1 Flash 1017 and GLM-5.3 on every test where they appear: on Terminal-Bench 4.0, 65.7 against 42.3 and 41.8. Claude Fable 5.1 beats it on three: Terminal-Bench Science, Frontier Finance and OSWorld.

Moonshot ran K3.1 Pro at max reasoning effort, temperature 1.0 and top-p 0.95, mostly inside Claude Code, with every model at its highest thinking setting. Most rival scores come from their makers' reports or public leaderboards; Moonshot ran some itself.

Moonshot's note says Claude models "may include fallbacks" and GPT models "may include cyberguards". Kimi Code Bench 3.0 is Moonshot's internal test.

Do independent tests back it up?

Mostly, yes. Artificial Analysis runs its own tests and ranks K3.1 Pro 4th of 215 models; among the rivals above, only Claude Opus 5.5 scores higher.

Artificial Analysis Intelligence Index v4.3.2, output speed and average cost per index task, checked October 24, 2026. Its 10 tests include Terminal-Bench 4.0, GDPval-AA and GDP.pdf, which are also on Moonshot's chart, but Artificial Analysis runs them itself.

Artificial Analysis also runs Terminal-Bench Science, outside its index. For these four tests, here are both results for K3.1 Pro:

The two land close: Artificial Analysis's runs came out higher on Terminal-Bench 4.0 and lower on GDP.pdf and Terminal-Bench Science. K3.1 Pro's 67.2% on Terminal-Bench 4.0 is also the best score in this group: Opus 5.5 got 59.6% and GPT-6 Astra 59.1% in the same tests.

Artificial Analysis's science run confirms the weak spot: 34.3% for K3.1 Pro against 59.0% for Opus 5.5 and 63.3% for GPT-6 Astra.

The catch is verbosity: K3.1 Pro wrote 150M tokens to finish the index, against a median of 87M, and Artificial Analysis calls it "very verbose". So a task costs about two-thirds of GPT-6 Astra's, though its price per token is a fifth.

Against Opus 5.5 it costs about a third as much per task, for a score 3 points lower.

Do the rival numbers hold up? Yes, where we could check them: Opus 5.5, GPT-6 Astra, Fable 5.1 and GLM-5.3 match their makers' reports or the public Terminal-Bench 4.0, SWE-Marathon and BenchCAD results. The GDPval-AA ratings match Artificial Analysis's live board to within 2 points.

Two caveats:

• Opus 5.5's 66.4 on Terminal-Bench 4.0 is Anthropic's own figure; the best public leaderboard run is 64.8.
• GPT-6 Astra's 75.8 on OSWorld comes from another company's report; OpenAI's own post gives 72.6 on a different OSWorld version.

What you can't check yet:

1. Most K3.1 Pro scores come only from Moonshot. Apart from the four above, the public boards have no K3.1 Pro result yet: the Terminal-Bench 4.0 leaderboard lists no Kimi model, and the SWE-Marathon board shows only Kimi K3, at 48.1.
2. Moonshot picks the rivals. The public SWE-Marathon board has Muse Wonder Full at 53.1, above K3.1 Pro's 51.7, and Muse Wonder isn't in Moonshot's chart.
3. Moonshot's pages disagree: the launch post puts the hallucination rate on its internal Hallu Bench Hard at 8.3%, down from 23.4% for K3, while the API guide says 6.85% and still shows unfilled placeholders such as "[xx.x%]". Use the launch post.
4. The weights come later: "by Nov 4", with no license named yet.

What changed from Kimi K3?

The price fell by a third and Moonshot's own tests went up; the way you call it didn't change.

All five are Moonshot's internal tests, from its launch post. Hallu Bench Hard is the share of false claims in its answers, so lower is better.

The API works like K3's: it always thinks, takes reasoning_effort as low, high or max (default max), and returns up to 131,072 tokens unless you raise max_completion_tokens. Temperature is fixed at 1.0 and top_p at 0.95, so leave them out; images go in as base64 or uploaded files, not public URLs.

Two things are new: the Highspeed version, and Moonshot retiring K2. K2 models close to new sign-ups on October 31, 2026, and Moonshot plans to retire every K2 model by the end of December.

In our Kimi K3 review, K3's always-on thinking at max effort made it 2 to 4.4 times slower than DeepSeek V4.1 Flash. One developer on Hacker News reports K3.1 Pro running "almost twice as fast as k3" on their own benchmark.

How much does Kimi K3.1 Pro cost?

$2 per million input tokens and $10 per million output: half of Claude Opus 5.5's price and a fifth of GPT-6 Astra's.

List prices from Moonshot's, Anthropic's and OpenAI's pricing pages, checked October 24, 2026.

A request with 10,000 input tokens and 2,000 output tokens costs $0.04 on K3.1 Pro, $0.06 on K3, $0.08 on Opus 5.5 or K3.1 Pro Highspeed, and $0.20 on GPT-6 Astra.

Real requests usually write more than that, because K3.1 Pro thinks before every answer, at max effort unless you lower it. Artificial Analysis's cost per task, above, includes the thinking.

Moonshot caches repeated prompt prefixes automatically: a cache hit costs $0.20 per million tokens, a tenth of the input price, and writing the cache costs $2.00 per million for 5 minutes or $4.00 for 1 hour. There are no price tiers by context length.

On Moonshot's API, K3.1 Pro unlocks after a top-up of at least $1. A new account starts at Tier 0 with 3 requests a minute and 1 at a time; $10 of total top-up lifts that to 100 a minute and 15 at once.

Should you switch to Kimi K3.1 Pro?

From K3, yes; from Claude Opus 5.5 or GPT-6 Astra, only after it handles a few of your own tasks.

If you run Opus 5.5 for coding, the reason to try K3.1 Pro is cost: about a third of Opus 5.5's cost per task on Artificial Analysis's suite, for an index score 3 points lower. Test it on your own repository first: apart from the two Terminal-Bench tests, its coding numbers are Moonshot's.

How do you start using Kimi K3.1 Pro?

Top up at least $1 on Moonshot's API platform and call kimi-k3.1-pro from any OpenAI-compatible client, or pick it in the Kimi app or Kimi Code.

This is the example from Moonshot's quickstart, with effort set to high:

import os
from openai import OpenAI

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

completion = client.chat.completions.create(
    model="kimi-k3.1-pro",
    reasoning_effort="high",  # low, high or max (the default)
    messages=[{"role": "user", "content": "Introduce Kimi K3.1 Pro in one sentence."}],
)
print(completion.choices[0].message.content)

1. In multi-turn chats and tool loops, send back the whole assistant message, including reasoning_content and tool_calls; Moonshot requires it.
2. Leave out temperature, top_p, n and the two penalties: they are fixed.
3. Send images as base64 or uploaded files; public image URLs don't work.
4. For Codex, Moonshot's guide sets model = "kimi-k3.1-pro" with wire_api = "responses" and the base URL https://api.moonshot.ai/v1.

OmniaKey doesn't serve K3.1 Pro as of October 24, 2026. It does serve Kimi K3 and every rival in Moonshot's chart, from Claude Opus 5.5 and GPT-6 Astra to DeepSeek V4.1 Flash and GLM-5.3, through one OpenAI-compatible API key. Run the same task on each and compare.

Common questions

Why can't my new Kimi API account use K3.1 Pro?

Moonshot unlocks it only after a top-up of at least $1. Your total top-up also sets your rate limits: Tier 0 allows 3 requests a minute.

Is K3.1 Pro Highspeed a different model?

Moonshot sells it as a version of K3.1 Pro that runs 2 to 3 times faster, at twice the price for input and output; cached input stays $0.20. OpenRouter lists both under one model, with Highspeed on its "priority" tier.

Can you turn off Kimi K3.1 Pro's thinking?

No. Moonshot's docs say it always thinks; set reasoning_effort to low to shorten it. OpenRouter's listing shows a "none" option, but Moonshot's docs list no such setting.

Is Kimi K3.1 Pro open source?

Moonshot says the full weights arrive by November 4, 2026; the API and apps work now. It also says it removed cyber-related training data from the open-weight version, and the launch post names no license yet.

How much does Kimi K3.1 Pro cost in China?

On Moonshot's China platform it costs ¥13 per million input tokens, ¥1.30 for cached input and ¥68 for output. Kimi K3 costs ¥20, ¥2 and ¥100.

Sources

• Moonshot: Introducing Kimi K3.1 Pro (October 23, 2026)
• Moonshot API: Kimi K3.1 Pro quickstart, reasoning effort, pricing and top-up and rate limits
• Moonshot China platform pricing
• Artificial Analysis: Kimi K3.1 Pro and GDPval-AA
• Terminal-Bench 4.0 leaderboard and SWE-Marathon leaderboard
• Anthropic: Claude Opus 5.5 system card and pricing
• OpenAI: GPT-6 Astra announcement and model page
• OpenRouter: Kimi K3.1 Pro
• Hacker News launch thread

All checked on October 24, 2026.