The Atomic Agent mascot holding a large dial marked low, high and xhigh, turned to xhigh, while a purple Qwen character points at a small price tag on a stack of coins
Models

Qwen 4 Max: Price, Benchmarks, and the Effort Setting Behind Alibaba's Numbers

Published
Reading time
13 min read
Tag

Qwen 4 Max passed all 59 hidden tests in our coding agent test for about $0.65 a task at high, slower than Opus 5.5. Price, benchmarks, effort levels.

Alibaba released Qwen 4 Max on November 16, 2026. Alibaba writes it Qwen4-Max. The API model ID is qwen4-max, and on OpenRouter it is qwen/qwen4-max. It costs $1.50 per million input tokens and $4.50 per million output, 25% less than Qwen 3.8 Max.

It is a large step up from Qwen 3.8 Max and still below the frontier. Artificial Analysis scores it 43 on its Intelligence Index, 9 points above Qwen 3.8 Max and 16 below Claude Opus 5.5 at max. Vals ranks it 11th of 38 and first among Chinese models, just ahead of Kimi K3.1 Pro. Artificial Analysis puts Kimi ahead, so the two evaluators disagree on that pair.

We also ran Qwen 4 Max ourselves on one coding task in Atomic Agent. At high, Alibaba’s default, it passed all 59 hidden tests in three runs, in about 25 minutes. It cost about $0.65 per task, the mean of the two fully recorded runs. Its predecessor, Qwen 3.8 Max, did not finish the same task in any of three runs in our completed 60-minute batch. One run in an earlier, aborted batch finished in 45 minutes. xhigh also passed everything, took about twice as long, and cost more. The other numbers come from Alibaba or outside evaluators, and we say which is which.


What Alibaba released

A preview called qwen4-max-preview opened on November 11 through Alibaba’s Token Plan subscription, according to ai-tldr.dev and AlphaSignal. Both report that the preview closes on November 25. The official blog post is dated November 16. Vals lists the release date as November 15, which fits the time difference between China and the US. OpenRouter routes qwen/qwen4-max to the snapshot qwen4-max-20261115.

According to Alibaba’s blog, the model has 2.4 trillion total parameters and 71 billion active per token. It also has 328 billion parameters in an n-gram embedding. Qwen 3.8 Max had the same total size and 95 billion active parameters, so less of the model runs on each token. Alibaba says the attention stack uses three Gated DeltaNet layers for each Qwen Sparse Attention layer. It also says the base model reaches 5 to 13% lower bits per byte than Qwen3.8-Max-Base, with 25% fewer active parameters.

Alibaba’s model documentation lists text, image, and video input with text output, tool calling including parallel calls, web search, structured output, and explicit and implicit caching. It works over the OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, and DashScope protocols. Thinking is on by default.

Alibaba has not published a knowledge cutoff or a license for the model. The blog lists QwenCloud, QwenWork, Qoder, and QwenStudio as places to use it.

Qwen 4 Max vs Qwen 3.8 Max

Qwen 4 MaxQwen 3.8 Max
Total / active parameters2.4T / 71B2.4T / 95B
Input / output, per 1M$1.50 / $4.50$2 / $6
Cache read on OpenRouter, per 1M$0.075$0.25
Artificial Analysis index4334
Cost per task on that index$3.27$10.02
Output tokens used on that index127M150M
Vals Index63.20%59.05%
Terminal-Bench 4.0, Alibaba-reported48.538.9
Open weightsNoYes, base 2.4T-A95B, custom license

Prices from Model Studio’s pricing page and OpenRouter, checked on November 16. Index figures from Artificial Analysis and Vals on November 16. Qwen 3.8 Max base weights are on Hugging Face under a custom license.

Per token, the new model is 25% cheaper. Per task on Artificial Analysis it is 67% cheaper, because it also used 15% fewer output tokens across the index. Alibaba’s own figure points the same way: on Work360, its internal benchmark, Alibaba reports a cost per task of $1.67 against $4.43 for Qwen 3.8 Max, and a score of 52.69 against 47.74.

The last row is a real change. Qwen 3.8 Max had open weights for its base 2.4T-A95B model, under a custom license. Qwen 4 Max has none.

The effort setting behind the headline numbers

Alibaba’s documentation lists three values for reasoning_effort: low, high, and xhigh. The default is high. Other common values are mapped silently:

You sendYou get
minimallow
mediumhigh
maxxhigh

From the Qwen 4 Max documentation, November 16. The Responses endpoint uses reasoning.effort with the same mapping. The Anthropic Messages endpoint uses output_config.effort and lists only medium to high and max to xhigh, with no minimal.

Two consequences follow. A client that sends medium to save money gets high and pays for it. A client that sends max gets xhigh, which is the most expensive setting.

Most of the published scores use xhigh:

  • Vals lists its run as xhigh, with temperature 1.0, top_p 0.95, and top_k 20.
  • Arena labels its WebDev entry “Qwen4-Max (xHigh)”.
  • Alibaba’s blog does not state the effort level in its benchmark footnotes. Judging by its token chart, the Terminal-Bench 4.0 result of 48.5 appears to be at xhigh.
  • Artificial Analysis does not say which level it tested.

Alibaba’s token chart also shows what the levels do to usage. On Work360 it reports 24% fewer tokens than Qwen 3.8 Max at xhigh and 35% fewer at high. On Terminal-Bench 4.0 it reports 48 to 49% fewer tokens. Alibaba has not published a cost per task at each level, and neither has any outside evaluator we checked.

So most of the scores describe xhigh, and the default API call runs high. The default thinking budget is 131,072 tokens. Until someone publishes high and xhigh side by side, start at the default and test xhigh on your own tasks before you pay for it.

What a task costs

Artificial Analysis runs a set of evaluations for its Intelligence Index and reports what each model spends. These are the figures it listed on November 16:

Model and settingIndexCost per taskTime per task
Fable 5.2, max64.6$26.5441.8 min
Opus 5.5, max59.1$10.0424.9 min
GPT-6.1 Sol, max54.6$1.2722.9 min
Kimi K3.1 Pro46.5$3.6840.1 min
Qwen 4 Max42.9$3.27not listed
Qwen 3.8 Max34$10.0248.0 min

Qwen 4 Max ranks 38th of 81 entries. Artificial Analysis does not state which effort level it used for Qwen 4 Max, while the Opus 5.5 score of 59 is at max. Opus 5.5 scores 52 at high and 48 at medium on the same index. Artificial Analysis updates these figures, and some differ from what it listed a week ago, so check the live page before you quote them.

Against Opus 5.5 at max, Qwen 4 Max scores 16 points lower and costs about a third as much per task. Against Kimi K3.1 Pro, it scores 3.6 points lower and costs 11% less on this index. On Vals and Arena, Kimi costs less per task, as the sections below explain. GPT-6.1 Sol at max scores 11.7 points higher and costs less than half as much, $1.27 against $3.27. On this index, Qwen 4 Max is cheap next to Claude and expensive next to GPT-6.1 Sol.

Artificial Analysis lists no time per task for Qwen 4 Max. For Qwen 3.8 Max it lists 48.0 minutes and 59 output tokens per second.

Arena’s Agent leaderboard gives a second cost figure. It ranks Qwen 4 Max 14th with a preliminary net improvement of 7.46%, a median task cost of $1.47, a median time of 15.7 minutes, and 47.1K output tokens per task. On the same board, Opus 5.5 at max has 18.19%, and Kimi K3.1 Pro at max has 3.95% in 21st place with a median task cost of $1.15.

What we measured ourselves

We ran Qwen 4 Max on one coding task through Atomic Agent, our open-source agent, on November 16. It is the task we used for Gemini 4 Carbon and Fable 5.2 vs Opus 5.5: a stock library of about 1,400 lines with 13 planted bugs and one missing feature, scored by 59 hidden tests. We used OpenRouter with qwen/qwen4-max, where Alibaba is the only provider, on a build included in the 0.7.0 release. Each setting ran three times with a 60-minute limit per run. All six runs ran in parallel. The other rows are from November 9, on the same task and build.

Model and settingHidden testsTimeStepsCost per task
Qwen 4 Max, high (Alibaba’s default; we set it explicitly)59/59, 3 runs25.0 min76about $0.65
Qwen 4 Max, xhigh59/59, 3 runs47 to 60 min75about $1 or more
Opus 5.5, medium59/59, 3 runs4.2 min10$0.97
Kimi K3.1 Pro, max59/59, 3 runs16 to 29 minabout $1.10
GPT-6.1 Sol59/59, 3 runs2.4 min19$0.29
Qwen 3.8 Max26, 58, 24 of 59stopped at 60 mindid not finish

At high, the runs took 24.1, 27.2, and 23.7 minutes. The cost is about $0.65, the mean of the two fully recorded runs, $0.62 and $0.68. The third and longest run was not fully recorded and likely cost more. At xhigh, two runs took 47.3 and 47.4 minutes. The 60-minute limit stopped the third, but its code already passed all 59 tests. The recorded xhigh costs were $0.84 to $1.28 per run, and each run is missing part of its requests, so the real cost is higher. The Kimi cost is from its one fully recorded run. Kimi’s three runs ran one after another on November 9, while the six Qwen 4 Max runs ran at once on one API key, which can make Qwen’s times look slower.

Qwen 4 Max finished the task in every run. Qwen 3.8 Max did not finish in any of three runs in our completed 60-minute batch on November 9. Those runs stopped at 26, 58, and 24 of 59 hidden tests, and the untouched code already passes 20. One Qwen 3.8 Max run in an earlier batch, aborted when our OpenRouter balance ran out, reached 59 of 59 in 45 minutes.

At the default high, Qwen 4 Max was the cheapest of the models that finished except GPT-6.1 Sol. It cost about a third less than Opus 5.5 at medium and about 40% less than Kimi K3.1 Pro at max. It was slow. It took about six times as long as Opus 5.5 and about ten times as long as GPT-6.1 Sol, and it worked in many small steps, 76 on average against 10 for Opus 5.5.

On our task, where high already passed everything, xhigh gained nothing. It roughly doubled the time and cost more. The reasoning token counts were 34K and 41K in the two fully recorded high runs, against at least 78K to 133K at xhigh. That suggests a higher setting reached Alibaba through OpenRouter.

Cache reads did happen. OpenRouter’s records show about 2.7 to 2.9 million input tokens per fully recorded high run billed as cache reads, even though OpenRouter lists implicit caching as off for this model. Atomic Agent sends no cache markers for Qwen, so this is Alibaba’s own caching. At $1.50 per million, that input would have cost over $4 per run, and each run cost $0.62 to $0.68.

This is one task and three runs per setting. Every finished run scored 59 of 59, so it says nothing about the ceiling of any model. The six Qwen 4 Max runs shared one machine and one API key. The comparison rows ran on a different day, and Kimi’s runs went one after another, so treat the time comparisons as rough.

The published speed figures for Qwen 4 Max do not agree with each other or with ours. Vals lists a latency of 64 minutes and 39 seconds at xhigh, and Arena Agent lists a median task time of 15.7 minutes. Artificial Analysis lists no time for it. These measure different work.

Independent results from Vals

Vals AI tested Qwen 4 Max at xhigh. Its figures on November 16:

BenchmarkQwen 4 MaxRank
Vals Index63.20%11 of 38
Vibe Code Bench86.66%15 of 129
Terminal-Bench 4.048.48%15 of 65
Finance Agent66.68%8 of 41

On the Vals Index, Qwen 4 Max is the highest Chinese model. Kimi K3.1 Pro has 62.52% and Qwen 3.8 Max has 59.05%. The top of the index is Fable 5.2 at 71.31%, about eight points higher, and Opus 5.5 has 70.14%, both as of November 16.

Vals lists Qwen 4 Max at $9.33 per test. On November 16 it listed Opus 5.5 at $24.77, Fable 5.2 at $17.94, and Kimi K3.1 Pro at $6.44, so on Vals Kimi costs less per test than Qwen 4 Max. These models do not all run at the same effort setting on Vals, so read those as what each entry cost in that run.

The Vals Terminal-Bench 4.0 score, 48.48%, matches Alibaba’s own 48.5. The other models’ figures do not match. Alibaba’s table gives Opus 5.5 59.6 and does not say where that number comes from, while Vals has Opus 5.5 at 65.2%. Alibaba gives Kimi K3.1 67.2, while Vals has Kimi K3.1 Pro at 63.6%. On Vals, Qwen 4 Max trails both by about 15 points on this benchmark.

Arena

On Code Arena WebDev, Qwen 4 Max at xhigh is sixth with a preliminary 1750 ±15 from 1,814 votes, as of November 16. Opus 5.5 at max has 1816, Kimi K3.1 Pro at max has 1686 in 13th place, and Qwen 3.8 Max has 1676. Preliminary ratings move as votes come in.

Arena and the Vals Index both rank Qwen 4 Max above Kimi K3.1 Pro. Artificial Analysis and the Vals Terminal-Bench 4.0 run put Kimi ahead. Cost per task splits the same way: Qwen 4 Max is cheaper on Artificial Analysis, and Kimi is cheaper on Vals ($6.44 against $9.33) and on Arena Agent ($1.15 against $1.47). If you are choosing between these two models for coding, the published data does not decide it.

Benchmarks from Alibaba

BenchmarkQwen 4 MaxQwen 3.8 MaxOpus 5.5GPT-6.1 SolKimi K3.1
DeepSWE v1.175.769.374.275.275.5
Terminal-Bench 4.048.538.959.656.167.2
ProgramBench31.528.065.052.525.0
CodeArena WebDev (Elo)17501676181617421686
Agents’ Last Exam46.727.943.442.643.4
APEX-Agents72.163.373.560.065.5
JobBench68.565.375.466.261.2
AutomationBench (AA)72.556.269.564.973.3
AA-Briefcase (Elo)16861605180115551674
HLAB-Vals21.315.49.26.311.7
AgentArena (net)7.5n/a18.211.83.9
HLE51.743.959.854.051.1

Compiled by Alibaba in the Qwen 4 Max blog post, November 16, 2026. Alibaba takes CodeArena WebDev and AgentArena from Arena, AA-Briefcase and AutomationBench from Artificial Analysis, HLAB-Vals from Vals, APEX-Agents from Mercor, and Agents’ Last Exam from its leaderboard. The other rows are Alibaba’s own runs. Alibaba labels the Kimi column “Kimi K3.1” and does not state effort levels in the footnotes. The table also includes GPT6-Astra, which we left out for width.

Alibaba’s own table does not show a clean win. Of the 12 rows, Qwen 4 Max has the top score on three: DeepSWE, Agents’ Last Exam, and HLAB-Vals. Opus 5.5 is top on seven, and Kimi K3.1 is top on Terminal-Bench and AutomationBench.

The coding rows split. On DeepSWE, Qwen 4 Max leads, but the four newest models are within 1.5 points of each other, and Qwen 3.8 Max is 6.4 points behind. On Terminal-Bench 4.0, it trails Opus 5.5 by 11 points and Kimi K3.1 by almost 19. On ProgramBench, Opus 5.5 scores more than twice as high, 65.0 against 31.5.

Pricing

Per 1M tokensQwen 4 MaxOpus 5.5Kimi K3.1 Pro
Input$1.50$4$2
Output, including thinking$4.50$20$10
Cache read$0.075$0.20$0.20
5-minute cache write$1.875$5$2
1-hour cache write$3$8not listed
Batchnot offeredhalf pricenot listed

Qwen 4 Max prices from Model Studio (International, Singapore) and OpenRouter on November 16. Opus 5.5 prices from Anthropic’s pricing page on November 9. Kimi K3.1 Pro prices from our Kimi K3.1 Pro review.

Per token, Qwen 4 Max input is about 2.7 times cheaper than Opus 5.5 and output about 4.4 times cheaper. Against Kimi K3.1 Pro it is 25% cheaper on input and 55% cheaper on output.

Five details affect the bill:

  • Model Studio charges the same rate up to 1M tokens, with or without thinking.
  • In the International (Singapore) region only, Model Studio gives 1 million free tokens. They are valid for 90 days from account activation, model release, or application approval, whichever is later.
  • Alibaba’s documentation lists batch inference as not supported.
  • In Alibaba’s Global and Beijing regions the price is $1.413 and $4.24.
  • OpenRouter has one provider for this model, Alibaba, and lists implicit caching as off for it. Alibaba’s caching docs say implicit cache hits are not guaranteed. Explicit cache breakpoints need a prefix of at least 1,024 tokens.

The caching line matters most in an agent. An agent loop resends the same long prefix at every step, and in our Gemini 4 Carbon test most input came from the cache, which kept the cost down. If cache reads do not happen on your route, you pay $1.50 per million instead of $0.075 for that input. In our test, Alibaba’s own caching applied through OpenRouter despite the implicit caching flag.

Model Studio’s rate limits are 30,000 requests per minute and 5 million tokens per minute in the US, Germany, Hong Kong, and Japan regions. In Singapore and Beijing the token limit is 5M, 10M, or 20M per minute, depending on monthly spend. AlphaSignal reports that the Token Plan subscription starts at about $3 for new users.

Limits to plan for in agents

  • Alibaba’s docs allow 131,072 output tokens without thinking and 262,144 for thinking and answer together. OpenRouter lists a maximum completion of 131,072. With a default thinking budget of 131,072, a long xhigh request on OpenRouter may reach that cap. We have not tested it.
  • The context window is 1,000,000 tokens. Input is limited to 983,616 tokens with thinking and 991,808 without. Artificial Analysis lists the context as 984k, which matches the thinking limit.
  • Alibaba’s docs list only low, high, and xhigh as effort values. Simon Willison wrote in the Hacker News thread that Qwen4-Max has a none setting. Alibaba’s docs do not list it.
  • Alibaba’s sampling defaults in thinking mode are temperature 1.0, top_p 0.95, and top_k 20.

Weights, smaller models, and fake repos

The weights of Qwen 4 Max are closed. Artificial Analysis lists the model as proprietary, and Vals lists its weights as private. Alibaba’s blog says “smaller models in the family are coming to open source” and gives no date.

On November 16, Model Studio’s pricing page listed only qwen4-max from the Qwen 4 family. Plus, Flash, 27B, and Coder models have not been released. The Qwen organization on Hugging Face had no new repositories after October 9.

Repositories with names like “Qwen4-27B-GGUF” have already appeared on Hugging Face from accounts that are not Alibaba’s. No Qwen 4 27B has been released, so these are not it. Download Qwen weights only from huggingface.co/Qwen.

Running Qwen 4 Max locally is not possible without the weights, and at 2.4 trillion parameters it would need a server with several data-center GPUs. llama.cpp has an open issue, #31072, to prepare for Qwen 4. If you want most of the work on your own machine, a smaller model is the practical choice today. We tested Qwen 3.8 27B as a local agent. Atomic Agent’s Fusion mode can also use a cloud model as the orchestrator, which plans and reviews, while local workers do the edits. We have not tried Qwen 4 Max in that role.

What early users report

The Hacker News thread on the launch had 534 points and 219 comments when we checked on November 16. The main topics:

  • Several commenters called the scores weak for a 2.4 trillion parameter model in late 2026.
  • Speed came up repeatedly. One commenter put Qwen 3.8 at about 30 tokens per second against about 80 for Opus, and others asked whether Qwen 4 Max is faster.
  • Many are waiting for the 27B and Flash models.
  • Some commenters raised censorship.

Hyperstack tested the preview at xhigh on November 11, on documents, vision, and video. It found Opus 5.5 ahead by 0.2 to 0.8 points, and Kimi ahead on three tasks. It did not test coding or agents, and it did not report cost.

Before our run, we did not find any published test of Qwen 4 Max inside a coding agent with cost and time per task.

How to try Qwen 4 Max

Download Atomic Agent for Mac, Windows or Linux, or install the terminal version. It is open source and works with your own API keys.

Terminal window
curl -fsSL https://atomicagent.io/install | sh
atomic-agent

On Windows, run irm https://atomicagent.io/install.ps1 | iex in PowerShell. The source is on GitHub.

Connect OpenRouter with your own key and choose qwen/qwen4-max from the live model list; see our OpenRouter guide. Atomic Agent sends no effort value by default, so Alibaba’s documented default, high, applies. In our test we set it explicitly. To try xhigh, add one line to your openrouter entry under llm.providers in ~/.atomic-agent/config.json:

"extraBody": { "reasoning": { "effort": "xhigh" } }

This is the line we used in our test, and the reasoning token counts suggest a higher setting reached Alibaba. On our task, where high already passed everything, it doubled the time and gained nothing, so try high first. The setting applies to every model on that provider entry, so change it when you switch models. More in our models guide.

If you are comparing Qwen 4 Max with other Chinese models, see our Kimi K3.1 Pro review. If you are comparing agents rather than models, see our list of open-source Claude Code alternatives.

The takeaway

Qwen 4 Max is a clear upgrade on Qwen 3.8 Max: 9 more points on Artificial Analysis for a third of the cost per task, and a lower token price. It is not a frontier model. It trails Opus 5.5 by 16 points on Artificial Analysis and 7 on Vals, and it trails by more on Terminal-Bench and ProgramBench in Alibaba’s own table. Against Kimi K3.1 Pro the evaluators split, on both score and cost per task.

In our one-task test it finished every run and passed all 59 hidden tests at both settings, while Qwen 3.8 Max did not finish in our completed 60-minute batch. At high it cost about $0.65 per task, the mean of the two fully recorded runs, less than Opus 5.5 and Kimi K3.1 Pro. It took about six times as long as Opus 5.5. On our task, where high already passed everything, xhigh doubled the time and gained nothing.

Remember that the default is high and most published scores are at xhigh, and that medium is billed as high. Start at high, and move to xhigh only after you measure a gain on your own work. The weights are closed this time, so for local use wait for the smaller models from huggingface.co/Qwen.

FAQ

Qwen 4 Max at a glance.

  • What is Qwen 4 Max?

    Qwen 4 Max, written Qwen4-Max by Alibaba, is the flagship model of Alibaba's Qwen 4 family. Alibaba's blog dates the release November 16, 2026, after a preview that opened on November 11. The API model ID is qwen4-max, and on OpenRouter it is qwen/qwen4-max. It has 2.4 trillion parameters with 71 billion active per token, a 1 million token context window, and accepts text, images, and video.

  • How much does Qwen 4 Max cost?

    On Alibaba Cloud Model Studio (International) and on OpenRouter it costs $1.50 per million input tokens and $4.50 per million output tokens. Cache reads cost $0.075, and cache writes cost $1.875 for 5 minutes or $3 for 1 hour. There is no batch discount. That is 25% less per token than Qwen 3.8 Max, which cost $2 and $6. In the International (Singapore) region, Model Studio gives 1 million free tokens valid for 90 days.

  • Is Qwen 4 Max open source? Can I run it locally?

    No. Artificial Analysis lists it as proprietary and Vals lists its weights as private. Alibaba says smaller models in the family are coming to open source, but on November 16 the Qwen organization on Hugging Face had no Qwen 4 repositories. Repos named Qwen4-27B from other accounts are not from Alibaba. Use only huggingface.co/Qwen.

  • Which reasoning effort should I use with Qwen 4 Max?

    The API accepts only low, high, and xhigh, and the default is high. Alibaba maps medium to high and max to xhigh, so sending medium does not save money. Vals and Arena tested at xhigh, and Alibaba's Terminal-Bench figure appears to be at xhigh. In our coding agent test, high passed all 59 hidden tests for about $0.65 per task, and xhigh also passed but took about twice as long and cost more. High had already passed everything on that task, so xhigh had nothing to gain there. Start at high.

  • How does Qwen 4 Max do in a real coding agent?

    We ran it in Atomic Agent on one coding task with 59 hidden tests, three runs per setting, through OpenRouter. At high, Alibaba's default, it passed all 59 tests in every run, in about 25 minutes and for about $0.65 per task, the mean of the two fully recorded runs. That is cheaper than Claude Opus 5.5 at medium ($0.97) and Kimi K3.1 Pro at max (about $1.10 in its one fully recorded run) on the same task, and about six times slower than Opus 5.5. Qwen 3.8 Max did not finish in any of three runs in our completed 60-minute batch. This is one task, so it says nothing about the ceiling of any model.

  • Is Qwen 4 Max better than Claude Opus 5.5?

    Not on independent scores. Artificial Analysis gives it 43 against 59 for Opus 5.5 at max and 52 at high, without stating Qwen's effort level, and Vals gives it 63.20% against 70.14%. It is much cheaper: $3.27 per task on Artificial Analysis against $10.04 for Opus 5.5 at max, and per token the input is about 2.7 times cheaper and the output about 4.4 times cheaper.

  • Is Qwen 4 Max better than Kimi K3.1 Pro?

    The evaluators disagree. Vals puts Qwen 4 Max slightly ahead, 63.20% to 62.52%, and Arena's Agent and WebDev boards also rank it higher. Artificial Analysis puts Kimi K3.1 Pro ahead, 46.5 to 42.9, and on Vals' Terminal-Bench 4.0 run Kimi leads 63.6% to 48.48%. Qwen 4 Max costs less per token. Per task it is cheaper on Artificial Analysis, $3.27 against $3.68, but Kimi is cheaper on Vals, $6.44 against $9.33, and on Arena Agent, $1.15 against $1.47.


Get new posts
by email

One email when we post something new.

Share
Written by
Nadya Dudka Product, Atomic Agent
Andrew Dyuzhov SEO and growth

Run your local agent
in one click