I currently have a set of agents that can run on the Kimi K3.1 Pro platform. When I saw the price list for Qwen4-Max today, my first thought was: with a cost of $1.5 to purchase and $4.5 to use, it’s significantly cheaper than Kimi’s options, which require $2 or $10 respectively – so definitely worth considering.

Before making the changes, there are at least four lines in the configuration file that need to be modified, two calculations that need to be redone, and one set of test results that require verification by yourself. The following explanation will follow the order in which these issues may arise during the migration process.

The value filled in for “reasoning_effort” in the first row represents the maximum value.

In the K3.1 Pro model, it is not possible to disable the reasoning function; the “reasoning_effort” setting only comes in three levels: low, high, and max, with the default value being “max”. Many users, for the sake of convenience, simply set it to “max” directly in the configuration settings.

The three settings of Qwen4-Max are “low”, “high”, and “xhigh”, with “high” being the default value. The “max” setting does not exist in its official documentation; therefore, it must be manually mapped to “xhigh” during the migration process. Otherwise, an error will occur, or the setting will be automatically adjusted back to the default level. In other words, even though you think the setting has been increased to its maximum, it may actually only be at half of that value.

The thinking modes on both sides are enabled by default. The difference lies in the way Qwen calculates the maximum output amount: when thinking is enabled, the total number of tokens that can be used for both thinking and generating the final answer does not exceed 262,144; the default budget allocated for thinking is 131,072 tokens. When thinking is disabled, the maximum single-output limit is still 131,072 tokens. According to Kimi’s documentation, the maximum output amount allowed is 1 million tokens.

Normally, no one would spit out hundreds of thousands of tokens at once. However, if your process includes a step such as “generating a lengthy report in one go,” the limitations that didn’t exist on Kimi might become obstacles when using Qwen.

Second line: temperature

Kimi locks the sampling settings at temperature 1.0 and top_p 0.95; attempting to use other values will result in errors. Therefore, code that runs successfully on Kimi most likely excludes these two parameters altogether.

After moving it to Qwen, it can be placed back there again.

For tasks that require stable results, such as data extraction, classification, or form filling, this feature is far more valuable than the price alone. In the five consistency tests conducted by Artificial Analysis using the AA-AnalystAgent, the K3.1 Pro ranked last among its four competitors. The most unstable responses were obtained when the same question was asked five times, and this was clearly due to the fixed sampling parameters used in the tests.

Third line: The context mentions 1M; should we believe it?

Both companies claim to have around one million pieces of context data. Both also include an interrogative mark in their statements.

The question mark in Kimi’s case represents an old issue: someone was running a long-duration task on the Bug Hunt Bench, causing the memory usage to peak at around 128K. Once K3.1 Pro started executing, it began producing inconsistent results, yet no error codes were reported by the system interfaces. This silent deterioration of performance could not be detected through the logs.

The question mark associated with Qwen represents a new development. According to Artificial Analysis, the context size designated for Qwen4-Max is 984K, not a full 1M; other media sources, based on observations of the preview version, claimed that its actual context size is 262K and only reaches 1M through additional expansion. Alibaba has not officially responded to this claim. If one of your reasons for migrating to Qwen is that it also supports a 1M context size and can be used directly, we recommend that you first test your longest tasks by running them ten times each at context sizes of 200K, 500K, and 900K. This step cannot be skipped or simplified in any way.

Fourth line: Interface format

This is undoubtedly the most hassle-free aspect of the migration process. Qwen4-Max is compatible with both OpenAI’s Chat Completions and Responses APIs, as well as Anthropic’s Messages API.

It turns out that when using the OpenAI-compatible format to configure Kimi, basically only the `base_url`, model name, and key need to be modified.

There is one thing to note: According to documents compiled by third parties, this system does not support fine-tuning or model deployment. As for whether batch inference is available, the information in these two sources contradicts each other. Teams that rely on the batch interface to save costs should check the console before going live.

Two price lists, and a transcript that doesn’t match any of them.

Kimi K3.1 Pro Qwen4-Max
Input/Output (dollars per million tokens) 2 / 10 1.5 / 4.5 (Singapore International Site)
Cache hit 0.20 0.075 (implicit caching)
output limit 1M 131,072; When engaging in thinking, the act of thinking itself is also included. 262,144
Reasoning Gear Shift Low / High / Max; the default value is Max. Low / High / XHigh; the default value is High.
Artificial Analysis Intelligent Index 3 points higher than Qwen4-Max 43
Cost per question for AA format 2.11 US dollars 3.27 US dollars
Terminal-Bench 4.0 67.2% 48.5 (Chart from Alibaba's press conference)

In terms of pricing, Qwen is significantly cheaper: when the discount is set at 25%, the final price drops by 55%; moreover, the cached prices offered by Qwen are less than 40% of those offered by Kimi.

However, the cost per individual item is calculated in the opposite way.

When Artificial Analysis completed the entire set of intelligent index questions, Qwen4-Max averaged $3.27 per question, while K3.1 Pro averaged $2.11 per question. The model whose price was half lower actually cost 55% more per question.

AA did not break down these specific numbers (3.27 dollars) in detail, but it provided two key pieces of information. The first relates to the output volume: During the entire evaluation process, the Qwen4-Max model generated a total of 127 million output tokens, while the median value for models in its category was 78.6 million. AA simply labeled this feature as “very verbose.” The second aspect concerns the input volume: The model consumed a total of 9.9 billion input tokens during the evaluation. In tasks involving intelligent agents, the system often reuses previous context data, but this detail is not visible in the price breakdown table.

It’s fair to add one thing here: Kimi’s score of 2.11 dollars was measured at the end of October, while Qwen’s score of 3.27 dollars was obtained today. The index version used for AA might have been updated in between, so these two numbers cannot be directly compared as if they came from the same test. However, such a significant difference in scores would not change just because there were minor updates to the measurement criteria.

The 60% mentioned by Alibaba, and the 10% measured by AA…

According to Alibaba, the average cost of completing a single task has decreased by 60% compared to the previous version. Specifically, on their own Work360 evaluation dataset, the cost dropped from $4.43 for Qwen3.8-Max to $1.67, representing a 62% reduction; at the same time, the amount of tokens required decreased by approximately 24% to 35%, and on Terminal-Bench 4.0, this reduction was nearly half. However, when measured by third-party organizations, the magnitude of this decrease was significantly smaller. According to Artificial Analysis, the intelligence score of this new version is 9 points higher than that of Qwen3.8-Max, but the cost per task has only decreased by about 11%.

Work360 is a evaluation dataset developed by Alibaba itself. It covers 26 primary categories and over 700 sub-specialties, and to this day, no external organization has replicated its testing process. I tend to interpret the 62% figure as the upper limit for performance when the cache capacity is fully utilized on what Alibaba deems typical tasks, while considering the 11% figure as a conservative estimate when making budget plans before any migration process.

Accounts should be calculated using conservative figures.

Those 20 points in the K3.1 Pro category… Qwen simply can’t make up for that deficit.

If your intelligent agent is primarily used in a terminal environment, where you continuously type commands, edit files, and run tests, then migrating it in this case is likely to result in losses.

On Terminal-Bench 4.0, the manufacturer reported a score of 65.7% for K3.1 Pro, while retests conducted by AA yielded a score of 67.2% – these two figures are consistent with each other. On Alibaba’s own press conference presentation chart, Qwen4-Max achieved a score of 48.5, whereas Claude Opus 5.5 scored 59.6 on the same chart. There is a difference of nearly 19 percentage points between these two scores, and these are indeed the figures that Alibaba chose to make public.

Alibaba’s strengths lie in other areas. Its self-assessed scores in DeepSWE v1.1 increased from 69.3 in the previous version to 75.7, while those in HealthBench Professional rose from 39.6 to 60.0. In the Image-to-WebDev Arena, it achieved a score of 1674, placing it sixth in the rankings. Areas where it made the most progress include generating web pages based on images, handling medical-related Q&A tasks, and resolving issues on GitHub.

The areas of expertise on both sides hardly overlap at all. Therefore, the question of “whether to switch or not” is not a holistic issue; rather, it needs to be answered separately depending on the type of task involved.

Those messy financial issues surrounding Kimi…

The motivation for making a transition often lies not in how much better the new model is, but in how unreliable the old one was. The few issues left behind by K3.1 Pro will no longer need to be addressed once you migrate to the new version.

When changing the reasoning level during multiple rounds of dialogue, the prefix cache becomes completely invalid. Many intelligent agents tend to downgrade when facing simpler steps and upgrade when encountering more difficult ones; however, on Kimi, such adjustments result in a loss of the 0.20-dollar discount and a reversion to the original price of 2 dollars with each switch. Additionally, some developers have noted that the token conversion recorded in the logs should have resulted in a deduction of 1.52 dollars, meaning the actual balance should have decreased by 2.02 dollars. Yet, the official team has not provided any explanation for this discrepancy.

In Kimi Code, if a single character in the model name is entered incorrectly, no error will be displayed; instead, an old model that lacks any cognitive capabilities will be used.

On AA-Omniscience, its illusion rate is 43.9%.

There are currently no publicly reported instances of these issues occurring with Qwen. However, Qwen4-Max was officially launched just today. The preview version has only been available within the Token Plan since November 11th, so it has only been a few days since users began using it on a large scale. The fact that there are no reported failures does not mean that such issues cannot occur; it simply means that no one has encountered them yet.

In the matter of open-source development, neither company can rely on it to achieve their goals.

There is a common belief on the internet: since all domestic models are open-source, choosing any of them will still allow for private deployment in case of need.

This alternative approach does not apply in either of these two models. The Qwen4-Max is closed-source, with only APIs available; Alibaba has stated that the smaller-sized models within the Qwen4 series will be gradually made open-source, but the Max model is not included in this plan. The K3.1 Pro is claimed to have 3 trillion parameters, while its predecessor, the K3, had 2.8 trillion parameters. To use the FP8 format, one would need an entire system equipped with 8 GPUs; moreover, the community-developed 1-bit quantization version of this model occupies a massive amount of storage space—570GB—in total, which is far beyond what most companies’ data centers can accommodate.

A team where data cannot leave the internal network is not a viable solution; neither of these options constitutes an appropriate answer.

Divide tasks among different groups instead of replacing everything in its entirety.

My approach is to route requests according to the type of task at the gateway level, rather than implementing a one-size-fits-all replacement strategy.

For terminal programming and for intelligent agents that require extensive independent operations in the command line, it is more appropriate to use K3.1 Pro. In such cases, fix the reasoning steps and set an upper limit for the output volume per operation. The difference of 19 percentage points represents a significant difference in success rates.

For tasks that involve creating front-end code based on screenshots, answering questions by reading documentation, or performing data extraction operations that require a stable format, switch to Qwen4-Max. Set the temperature to a lower level and try to keep the cache prefix fixed as much as possible. This way, you can make full use of the cached data at a price of only $0.075 per unit. It is in these areas that the pricing advantage of Qwen4-Max truly comes into play.

In the first week after implementing this change, I kept in mind AA’s reminder regarding the 127 million output tokens generated per request: I set a maximum output limit for each individual request and checked the costs on a daily basis to see whether they actually fell within the range of $1.67 or $3.27 per request.

The bill for this week is a better indication of whether it’s worth the effort than any percentage figure mentioned at a press conference.

In-site entry options: Kimi, Tongyi Qianwen, Qoder.

Relevant comparisons: “Kimi K3.1 Pro vs. DeepSeek V4.1 Flash 1017: The difference in score for each question is eight times; the extra money spent only bought the ability to perform terminal programming tasks.” “Kimi K3.1 Pro, GPT-6 Astra, GPT-6.1 Sol: The K3.1 Pro, which has exactly the same price as the Sol version, costs 2.9 times more per question.” “LongCat-2.5, GLM-5.3, DeepSeek V4.1 Flash: On Meituan’s report card with 31 rows of results, our own column is filled solely with self-test scores.”