Kimi K3.1 Pro hands-on test: Why does quota consumption still spark controversy after the API price cut?

By | Shuzhi Benliu, Author | Tian Yu, Editor | Shen Xiao

In the early hours of October 24, Kimi released its new model K3.1 Pro. The official announcement set out two main upgrades: its capabilities have entered the first tier of flagship models, and API prices have been cut by about one-third compared with the previous generation.

But after the model's release, Kimi got caught up in a cost-effectiveness controversy because of how fast it consumes Tokens.

Looking at the pricing, the officially announced API prices did indeed fall.

K3.1 Pro's input price dropped from 20 yuan to 13 yuan per million Tokens, and its output price from 100 yuan to 68 yuan. Yet for consumer users and overseas developers, the actual experience did not get cheaper in step.

Some users tested the classic "pelican riding a bicycle" task; generation took 1 hour and 47 minutes, consumed 87% of the 5-hour quota and 16% of the weekly quota. Another developer recorded that sending just one sentence, "Hello," pushed the Token count to 42,000.

On October 26, Kimi issued an apology and compensation statement, admitting that "the problem was ours, and we should have fixed it sooner." The company proposed two measures: resetting and compensating the quotas of subscription members within their validity period; and adjusting the model configuration to reduce unnecessary sub-Agent calls and output length.

The statement acknowledged that the invocation mechanisms behind the model had not been properly controlled.

In the past, users understood model pricing relatively simply: however many Tokens are output, the unit price is applied.

But now, a single request may also include long-context reading, background reasoning, task decomposition, tool calls, sub-Agent collaboration and multiple rounds of self-correction.

A user sends only one sentence, but the model may complete a whole string of actions in the background. As long as each task consumes more Tokens, even if API prices fall, the cost users ultimately pay may still rise.

K3.1 Pro is officially described as "the most capable 3-trillion-parameter model to date," supports a 1-million-Token context, and is positioned for long-horizon programming, 3D creation and financial research. In these strength scenarios, it is inherently prone to triggering task types that involve long contexts and multi-Agent calls.

In tests by the independent research organization Artificial Analysis, K3.1 Pro scored 55 on the intelligence index, ranking fourth globally and first among open-source models, 11 points higher than K3. On the strength of this ranking, if the model weights are subsequently opened as planned, it is also expected to become the highest-scoring open-weight model on that list.

These results show that Kimi has indeed completed a leap in capability.

But between base-model performance and product experience, there is still ordinary users' real experience of using it.

To observe how K3.1 Pro performs in everyday tasks, we used a Pro subscription account and ran a series of creative tasks through the Kimi desktop app.

The focus of the tests was not whether a single answer looked good, but four stages: the first-generation result, the revision process, the final degree of completion, and quota consumption.

As of the afternoon of October 27, the account's total monthly usage was 27.03%, of which the five test tasks together consumed about 25.48%.

A pelican riding a bicycle is a classic hands-on test that appears after almost every new model release. Its origin is the tech blogger Simon Willison. The original prompt was just one sentence: Generate an SVG of a pelican riding a bicycle.

An SVG requires describing shapes in code, so the model must understand both the structure of the pelican and the bicycle, and handle the spatial relationships among the wheels, frame, bird legs and pedals. Why does Simon keep picking on "a pelican riding a bicycle"? First, because Simon himself likes it; second, because similar material was scarce online at the time, making it hard for models to directly reproduce images from their training data.

As model capabilities have advanced, this test has gradually become a "boot-up ceremony" after each new model release. This time we upgraded the prompt to a short video, to keep observing, beyond structure and aesthetics, whether the riding motion is natural and whether the visuals are coherent.

We asked for a smiling pelican riding a bicycle along the coastline at sunset, with waves constantly lapping the shore.

The task was completed in one pass with no follow-up revisions, and the final result was good: the pelican's expression was vivid, its feathers lifted in the wind; the waves lapped naturally, and the overall style was warm.

The quota breakdown showed that the K3.1 Pro model entry consumed 0.13% and the Plugin entry consumed 5.85%, for a total of 5.98%. In other words, this task involved no rework, and about 97.8% of the consumption came from the Plugin entry.

Next, we had it create an old Suzhou-style courtyard.

The prompt was "make an old Suzhou-style courtyard, with a loquat tree in the yard, a rattan bench and a table under the tree, and loquats and a teacup on the table."

The first version was rather rough, with an overall texture close to grain building blocks and leaves like pieces of paper.

For the second version, we asked it to drop the sandbox-style top-down view in favor of natural perspective and slight depth of field inside the courtyard; use low-angle warm sunlight to strengthen the summer atmosphere; replace the bench with a lounge chair; and continue refining other details.

This version solved the main problems of viewpoint, lounge chair and palm-leaf fan, but details still deviated. The bamboo grove outside the courtyard was overly simplified, the palm-leaf fan floated in midair, and the loquats were embedded in the tree roots.

For the third version, we provided a real reference image and had it redo the scene. With the reference image, the content became richer, and the problems concentrated on detail fidelity. The stone pillars were too prominent, and the stones around the pond had a texture close to foam.

The quota breakdown shows that the Su-style old courtyard consumed 1.76% on its first generation, subsequent revisions added another 4.45 percentage points, and the entire task consumed 6.21% in total, ranking second among the five tests.

The third task was to create a webpage introducing the history of world art. We required a sufficiently rich set of representative figures and paintings, so that when users click on a figure's avatar or a painting, they can view related knowledge explanations.

The generated webpage was named "ATHÈNE | World Art History · Romantic Corridor," using warm paper tones, deep browns and soft light and shadow to create a romantic atmosphere. Users can filter by cultural track, figure or painting, roam by chapter, search keywords, and awaken figure animations.

The first version was closer to a curated gallery, containing only 13 figures and 14 masterpieces—rather few—and Chinese and Western art were mixed into a single thread. When clicking "Guide to the next piece," the text explanation on the right and the painting on the left were out of sync; the painting appeared about three seconds later than the text.

After further optimization, the webpage content expanded to 30 figures and 30 works, presented in dual tracks of Western art and East Asian art, with a timeline clearly organizing the context. Clicking any figure's avatar, painting or timeline node lets users read the relevant background introduction. It also browses smoothly on mobile, with details balancing aesthetics and usability.

The art history webpage's model entries consumed 4.90% in total, and Plugin entries consumed 0.02%, for a total of 4.92%.

Next, we reproduced the tests of two bloggers.

The first was a mechanical keyboard; we asked it to create a high-precision model showing the structure of parts such as keycaps, switches, the positioning plate, PCB and casing, while supporting white model, X-ray perspective, exploded view and different directional viewpoints.

This generation did not go through multiple rounds of redoing; the main structure was complete, and the various views could be switched normally. Two billing records appeared in the quota breakdown, at 1.09% and 0.37% respectively, totaling 1.46%, the lowest among the five tasks.

The second task was to recreate "The Peach Blossom Spring."

The first version's imagery was rather symbolic, with figures and objects replaced by abstract elements, and the overall texture was close to the first version of the Su-style old courtyard. We then provided a landscape painting reference image and had it redo it. The second version still failed to capture the spatial changes and narrative layers of the original text, and the imagery was clearly distant from the mood of "The Peach Blossom Spring."

We made one last attempt, sending it the original blogger's generated video and asking it to reproduce it as closely as possible. This version turned out much better, with clear improvements in overall texture and image completeness. There were also a few places where problems were obvious at a glance, such as the light that appeared multiple times in the image looking like hollow columns, lacking a realistic lighting effect.

This task's quota breakdown had nine records in total, consuming 6.91% cumulatively, the highest among the five tests. The last attempt and the subsequent adjustments alone consumed 2.88%.

Putting these two cases together, the mechanical keyboard was completed in one go, consuming only 1.46%; "The Peach Blossom Spring" went through multiple rounds of revision and two rounds of reference material provision, consuming 6.91% cumulatively, nearly five times the former.

Returning to the product itself, overall, the five tasks consumed about 25.48% of the monthly quota in total. For complex creative tasks that require repeated revisions and tool calls to complete, this level of consumption is already not low.

Among them, "The Peach Blossom Spring" consumed 6.91%, the Su-style old courtyard 6.21%, the pelican video 5.98%, the art history webpage 4.92%, and the mechanical keyboard 1.46%. Among these, "The Peach Blossom Spring" and the Su-style old courtyard still had room for improvement in their final results after multiple rounds of attempts.

This set of records reveals three types of cost: the cost brought by repeated revisions, the cost brought by tool calls, and the extra consumption related to reference materials and repeated generation.

Beyond the product experience, Kimi's recent overseas promotion controversy has also been considerable.

Kimi found an overseas blogger to do a promotion, but asked them to hide the advertising label and pretend it was the blogger's own genuine experience shared voluntarily.

Someone commented: "Kimi K3.1 Pro itself is a very good model; there was no need to recruit people to post fake positive reviews, but they insisted on cutting corners and playing tricks, making the whole product launch look terrible."

According to him, creators participating in the campaign had to submit copy in advance, get it reviewed, and were asked not to add the "Paid Partnership" paid collaboration label; the brief also contained promotional language comparing Kimi with Claude Opus 5.5.

However, the relevant documents have not yet been verified, and it cannot be confirmed whether they were issued by Moonshot AI or a third-party agency.

Several overseas bloggers also reported that after posting content related to Kimi K3.1 Pro, they were penalized by the X platform.

Review blogger Lumina said that they had only tested the model through Arena before its release, neither receiving payment nor obtaining early access from Kimi officially, yet two of their posts were penalized.

According to these bloggers, some accounts had to add a label first before they could restore their accounts or file an appeal. Some, in order to get their accounts back, could only add the label as required and then delete the related posts, so that fans would not mistakenly think they had been paid for promotion.

X platform rules require that natural content belonging to paid collaborations must display the "Paid Partnership" label; the platform has not yet publicly explained why these accounts were handled in a concentrated manner.

For the outside world, this adds new doubts to an already controversial product launch: between paid promotion, independent reviews and platform moderation, what exactly happened?

In June this year, former miHoYo international president Jin Wenyi joined Moonshot AI, responsible for Kimi's overseas expansion and globalization, and will be based overseas in the future. According to reports, Jin Wenyi left miHoYo at the end of May and joined Moonshot AI after completing the handover.

At miHoYo, Jin Wenyi had long been responsible for overseas publishing and international business, participated in building the HoYoverse global brand, and experienced the process of products such as "Genshin Impact" and "Honkai: Star Rail" going from China to the world.

Her joining Kimi shows that the problem Moonshot AI needs to solve has changed: Kimi no longer only needs to make stronger models, but also needs to deliver models to users in different countries and build a global business that can operate continuously.

At the beginning of this year, Moonshot AI mentioned in communications with investors that Kimi's overseas revenue had already exceeded domestic revenue; after the release of K2.5, global paying users grew fourfold.

Kimi's overseas growth comes from excellent model performance, but the globalization of a Chinese large-model company will inevitably also involve dealing with localization, channels, branding and trust issues.