Kimi K3.1 Pro: The Strongest Open Model Yet, but Not the Cheap Opus You're Being Promised

October 23, 2026, a few hours after launch

The feed fills up with videos: a rocket lifting off, a rooftop pool at sunset, a robotic arm breaking apart into 2,317 pieces. All of them were built with Kimi K3.1 Pro, and many repeat the same message: this is Opus 5.5, just open source and several times cheaper.

That evening, a user on X posts a screenshot of what looks like a Moonshot brief for influencers. One line reads: "Please DO NOT put on the paid partnership label."

The model really is good. It's just hard to tell which videos were tests and which were ads.

Kimi K3.1 Pro is the new flagship model from Moonshot AI, a Chinese AI company, released on October 23, 2026. I went through the official numbers, the one independent measurement available and dozens of developer tests, and filtered out the ones that look like part of the campaign. I haven't run the model on client tasks yet. One term that comes up throughout: Moonshot promises to publish the "weights" by November 4, meaning the model file itself, so other companies can run it too.

My take, at a glance

#4The strongest open model yet55 points on the independent index, vs. 58 for Opus 5.5

Up to 20xCheaper than Opus on visual demos3.5x to 20x on the same task. In most tests Opus still looks better

36Behind on real workOn a 105-bug test: Opus 41.7, DeepSeek 42.7

WaitWhat it means for a businessUntil November 4 at least, when the model should open up and more companies will run it

My take: a serious contender, not a replacement

This is proof that an open model can reach fourth place in the world. The next open models scored 45. Kimi scored 55, just three points below the leader. Even people who never use it will benefit: when a model like this is open for download, providers compete on price, and the pressure on OpenAI, Anthropic and Google to cut prices grows.

Where Kimi stands: the Artificial Analysis index

Lower scoreThe leader

58: Claude Opus 5.5

56: Claude Sonnet 5.5, Muse Wonder Full

55: Kimi K3.1 Pro

53: Gemini 4 Argon, GPT-6 Astra, Claude Fable 5.1

52: GPT-6.1 Sol

45: DeepSeek 1017 and GLM-5.3, the next open models

Source: Artificial Analysis, October 24, 2026. Kimi ranks fourth out of 215 models.

What the feed said, and what was measured

"Open-source Opus 5.5"Three points below Opus on the index, and behind it in most tests. The exception: working on its own at a computer

"5x cheaper"Half Opus's price per token, and about a third per task

"Open source"Not yet. The weights are promised by November 4

And if you run an agent, there's another problem: GPT-6.1 Sol charges exactly the same per token, and finishes a task at a third of the cost.

What Kimi K3.1 Pro is, in a minute

It's July's Kimi K3 after another big round of training, not a new model built from scratch. According to Moonshot, it's especially strong at building complete coding projects, 3D, and financial tasks.

$2per million tokens it reads, and $10 per million it writes. Half of Opus 5.5

1Mtokens it can read at once, with no surcharge for long text

×2A version 2 to 3 times faster, at double the price

PriciestThe default: it thinks as much as possible before every answer. You can turn it down, not off

Sources: Moonshot's pricing page and launch blog, as of October 24, 2026. We explain tokens in What is a token.

The independent measurement: smart, but not cheap per task

Artificial Analysis, a company that measures models independently, ran it on launch day. It's the only major measurement available right now. The score is excellent, but the cost per task tells a different story:

What an average task costs, and the score

Claude Opus 5.5 · score 58

$5.98

Kimi K3.1 Pro · score 55

$2.11

Gemini 4 Argon · score 53

$1.99

Opus 5.5, reduced reasoning · score 54

$1.82

GPT-6.1 Sol · score 52

$0.72

DeepSeek 1017 · score 45

$0.26

Source: Artificial Analysis, October 24, 2026. Each model at its highest reasoning setting, except Argon and the Opus reduced-reasoning row.

Three numbers that explain the chart

55KWrites a lot, so it's pricey55,000 tokens per task, vs. 38,000 for Sol. You pay for every one

67%Working on its own at a computerTasks where it runs computer commands by itself. Here it leads: Opus scored 60%

30Factual knowledgevs. 42 to 46 for Opus, Argon and Sol. It makes things up more

Source: Artificial Analysis. Why a model making things up matters for a business: What is an AI hallucination.

Same task, several models: what actually came out

I left out every test that repeats the wording from the brief, even if some may be genuine. What's left are testers who show losses too, or who run their own benchmarks.

Five tests, and who won according to the tester

Mentos in CokeKimi: $0.71, 5 minutes. Opus: $2.94, 16 minutes. Opus more impressive, Kimi better value

A 3D sceneKimi: about $7.30. GPT-6 Astra: about $86.50, in the same time. Kimi won

Four space scenesKimi: $1.09. Opus: $5.63. Opus won, Kimi "flatter"

A castle above the cloudsKimi: 44 minutes. Opus: 35 minutes. Opus "still unbeatable"

105 real bugsKimi: 36. Opus: 41.7. DeepSeek 1017: 42.7

Green: Kimi won. Red: it lost. Orange: depends on what you measure. Source: the testers' posts on X, October 23–24, 2026: ottollm, notjazii, hug0_plat, sakatayasha and Pawel Huryn.

Mentos in Coke: less impressive, much cheaper. Opus built a 5.6-meter geyser for $2.94. Kimi built a 4.4-meter geyser for 71 cents, in a third of the time.

ottollm's post and comparison video on X, October 24, 2026

In his view, Kimi is the best value and Opus the most impressive. His comparison, not ours.

Against GPT-6 Astra: $7 vs. $86. Same working time, 12 times the price. The tester, who last week also showed DeepSeek's losses, gave this one to Kimi.

notjazii on X: GPT-6 Astra vs. Kimi K3.1 Pro, October 24, 2026

Both models at the highest reasoning setting.

Against Opus: 5x cheaper, "but the quality isn't there yet." Opus's scenes looked like they came from a movie. Kimi's came out flat, and the galaxies got swallowed by the dark.

hug0_plat on X: four space scenes, October 23, 2026

The tester preferred Opus.

Against Opus, again: "still unbeatable." A castle above the clouds. Kimi did a good job, but Opus was faster and prettier.

sakatayasha on X: a castle above the clouds, October 23, 2026

"Kimi honestly did a solid job… but imo, Opus is still unbeatable."

A tester who wrote "not sponsored." ChrisGPT tested it a week before launch. In the trailer he built, Kimi came out slightly prettier than Opus. For regular work he still prefers Opus: "Kimi k3.1 is not Opus 5.5 tier please don't be ridiculous." And another developer, Ken Gray, ran the prompt from one of the viral videos, and for him Opus actually won.

ChrisGPT on X: a trailer built with Kimi K3.1 Pro, October 23, 2026

The visuals are good, but for regular work he prefers Opus 5.5.

Where it lost

Real bugs: behind DeepSeek too. We first saw Pawel Huryn's test when we covered DeepSeek 1017: 105 real bugs across two projects. It's the task closest to what an agent does on a business's website. And Kimi's results jump around: across three runs it scored 30, 36 and 42. A model like that is hard to trust without checking its work.

105 real bugs: average score over three runs

DeepSeek 1017

42.7

Claude Opus 5.5

41.7

Kimi K3.1 Pro

Source: Pawel Huryn on X, October 23, 2026, all models at the highest reasoning setting. Opus scored 39, 43 and 43 across its three runs. DeepSeek's result comes from Huryn's updated table, so it differs from the one in our earlier article.

Pawel Huryn on X: 105 real bugs, October 23, 2026

"Considering the cost, I see no reason to use this model."

LiveBench: below DeepSeek. LiveBench is a benchmark that keeps changing its questions, so no one can prepare for it in advance. Bindu Reddy, who runs it, wrote that Kimi scored lower than DeepSeek Flash. In her view, it shines mainly on the benchmarks the company chose to show.

Bindu Reddy on X: Kimi K3.1 Pro on LiveBench, October 23, 2026

"Not Fable or Astra class." In her view, DeepSeek is still the best open model.

The campaign: what's known and what isn't

elshayib_ on X: a screenshot of an influencer brief, October 23, 2026

"Please DO NOT put on the paid partnership label." Moonshot had not responded by the time we published, and we have not verified the document ourselves.

What's known

Screenshots of a brief, posted on X
The request not to label sponsored content
Ready-made lines, like "Open-source Opus 5.5, 5x cheaper"
The same lines, word for word, in several of the most-viewed posts

What isn't known

Who wrote the document, and whether it's real
Who received it
Which posters were paid
A response from Moonshot

Even people who say they never got the brief were angry. A well-known AI news account on X wrote that Moonshot "ruined their launch themselves," and that the model didn't need people faking good reviews.

Three questions before you believe an "X kills Opus" video

Any losses?An honest tester also shows a task where the model failed

Time and cost?"Amazing" without numbers isn't a test

Is it like your business?A business agent works on email, spreadsheets, the website and customers, not rocket launches

What it means for a business with an agent

✓

Worth trying nowIf this is your business

Lots of animated visual content
Animated landing pages
3D product renders
Models for 3D printing

⏳

Wait for November 4And what comes after

The open weights
More companies running it
Lower prices from them
More independent measurements

Don't switchToday

Agent on GPT-6.1 Sol: it'll cost more
Website and code fixes: Opus and DeepSeek do better
Customers' personal data

If your agent runs on Opus or Sonnet at the highest reasoning setting, Kimi is 2.5 to 3 times cheaper per task. But on non-visual work it's likely to make more mistakes, and Opus with reduced reasoning already costs less than it does. And for really big savings, DeepSeek 1017 did better on both non-visual tests, at roughly an eighth of the price.

Where the data is stored. Moonshot's service for developers and agents is operated from Singapore, and the data is stored there. The terms allow Moonshot to use your content to improve the service, unless you've signed a different agreement. So I wouldn't send it customers' personal data. We covered what Israeli law requires in AI and Amendment 13; if you're outside Israel, check what your local privacy rules require.

What changes after November 4. Here's what happened with K3, the previous version, after it was opened:

Large companies like Fireworks, Together and Amazon Bedrock started running it.
Their prices started at $0.19 per million tokens, instead of $3 at Moonshot.
Fireworks announced it runs the model in the US, without storing the data.

If that happens again, you'll be able to use Kimi without sending data to Moonshot.

What I'm doing about it

I usually build agents on Hermes Agent, an open-source framework from Nous Research, and choose the model that fits each business: OpenAI, Anthropic, Google, DeepSeek, Moonshot or others. Platin Digital is an OpenAI Select Partner, and that doesn't change the choice.

Kimi K3.1 Pro isn't going into my client work for now. Once the model is open, another company runs it and more independent measurements come in, I'll test it myself on visual content and update here.

In our ranking, the Sigma AI Ranking, K3.1 Pro appears in the "New in the ranking" strip, and will get a score and a place once the independent measurements arrive.

For comparison: Gemini 4 Argon and Anthropic's comeback with Opus 5.5. And what an agent costs beyond the model's price: How much an AI agent costs a business.

Frequently asked questions

What is Kimi K3.1 Pro?

Kimi K3.1 Pro is the flagship model from Moonshot AI, a Chinese AI company, released on October 23, 2026. It's a version of Kimi K3 with a large round of additional training. On the independent Artificial Analysis index it ranks fourth, with 55 points, the highest score yet for a model that's about to be open. Moonshot promises to publish the weights by November 4, 2026.

How much does Kimi K3.1 Pro cost?

On Moonshot's API: $2 per million input tokens, $10 for output, and $0.20 for text the model has already read. The Highspeed version, 2 to 3 times faster, costs double. There's no surcharge for long text and no peak hours. On the Artificial Analysis index, an average task cost $2.11, vs. $5.98 for Opus 5.5 and $0.72 for GPT-6.1 Sol.

Is Kimi K3.1 Pro as good as Claude Opus 5.5?

No. On the independent index Opus 5.5 scored 58 and Kimi 55. In most of the developers' visual tests Opus came out more polished, and on the 105-real-bug test Opus scored 41.7 vs. 36 for Kimi. Kimi's advantage is price: on 3D renders it cost between 3.5 and 20 times less in the tests.

Is Kimi K3.1 Pro open source?

Not yet. Today it runs only on Moonshot's servers. The company wrote that it will publish the full weights by November 4, 2026. The K3.1 license hasn't been published yet. K3's license was an MIT-style license, with extra conditions for very large companies.

Is it safe to send business data to Kimi?

According to Moonshot's privacy policy, what you send to the API is stored on servers in Singapore, and the terms of use allow it to be used to improve the service, unless there's a different written agreement. For content that goes public on your website anyway, that's not a problem. For customers' personal data, it's better to wait for hosting providers outside Moonshot once the weights are out, or to choose a different model.

What's the story with the influencers and sponsored content?

On October 23, a user on X posted screenshots of what looks like a Moonshot brief for influencers in the launch campaign. In the brief, influencers were asked not to label their posts as paid partnerships, and were offered lines like "Open-source Opus 5.5, 5x cheaper." Moonshot had not responded at the time of writing, and we have not verified the document ourselves. It's unknown which posters were paid.