New Chinese Kimi K3.1 Pro AI Model Outperforms GPT-6 Astra With Near Claude Opus 5.5 Performance

Chinese AI developer Moonshot AI has released Kimi K3.1 Pro, a model whose biggest gains come from carrying out sequences of coding and business-software tasks. Available through Kimi’s apps and developer API since October 23, the upgrade gives teams building AI agents a stronger option for work that requires repeated tool use, debugging and revision. The API lets other applications use the model remotely.

Independent evaluator Artificial Analysis measured a jump from 13% for Kimi K3 to 67% for K3.1 Pro on Terminal-Bench 4.0, a test of coding agents performing terminal tasks. Its business-software workflow test rose from 58% to 73%. Smaller improvements on knowledge and science tests make this primarily a gain in sustained task execution.

The release also faces allegations about undisclosed influencer promotion.

The Largest Gains Come From Repeated Tool Use

K3.1 Pro builds on the Kimi K3 generation introduced in July. It retains native image input and a one-million-token context window, the input a model can consider in one request. Moonshot attributes the improvement to substantially more post-training, including reinforcement learning across many task environments.

That training targets work in which a model chooses an action, receives a result and decides what to try next. The surrounding agent software supplies tools and an execution environment. A useful answer can therefore require many rounds of work rather than one direct response to a prompt.

Artificial Analysis’s October 23 evaluation puts the change in perspective:

K3.1 Pro reached 55 points on the evaluator’s Intelligence Index, up from K3’s 44. At launch, that placed it fourth, behind Anthropic’s Claude Opus 5.5 at 58 and Claude Sonnet 5.5 and Meta’s Muse Wonder Full at 56.

The aggregate brings together different evaluations; the much larger terminal and workflow gains explain more about the upgrade’s practical emphasis than its position alone.

Moonshot’s walking-simulator demonstration shows the sort of diagnostic loop it is trying to improve. In an internal project built on human-atlas, an anatomical model of the body, the company says Kimi repeatedly evaluated and revised a walking controller, taking the simulated body from collapse to 11.2 seconds of alternating strides. When changes to the training reward failed to improve arm movement, it investigated the simulator itself, corrected 36 errors in calculations governing shoulder-muscle leverage and retrained the controller.

Where Other Models Remain Competitive

Artificial Analysis’s Kimi–Opus comparison, reviewed October 25, shows why the task matters. Kimi leads on Terminal-Bench, 67% to 60%, and AutomationBench, 73% to 70%. Opus leads on Humanity’s Last Exam, 61% to 52%, and SciCode, 67% to 61%. These are the evaluator’s configurations, including Opus at maximum effort with its default fallback.

Moonshot’s own Terminal-Bench comparison instead puts Kimi at 65.7%, just behind Opus at 66.4%. Its comparison assembles results from different reported runs and agent software.

Other choices remain meaningful. Claude Sonnet 5.5 scores one index point above Kimi and 64% on the terminal test. DeepSeek V4.1 Flash 1017 trails Kimi on that test, 34% to 67%, but already offers weights for download and much lower token prices.

Xiaomi’s MiMo-V2.6-Pro has an index score of 46, yet it matches Kimi’s 61% on SciCode and exceeds Kimi on CritPt physics problems, 27% to 25%. Both pages use Intelligence Index version 4.3.2. MiMo’s lower overall score leaves room for a cheaper model to remain competitive on particular work.

From 3D Scripts to Financial Documents

Moonshot also develops the model’s practical role beyond conventional code. It says K3.1 Pro can write and execute Blender and CAD scripts to change geometry, materials and scene layouts. Its 3D Make feature lets users preview, edit and export models through Kimi’s web and mobile products. In Moonshot’s BenchCAD evaluation, the model reconstructs shapes from multiple rendered views by generating CAD code.

For financial work, Moonshot’s OfficeQA Pro v2 evaluation requires finding information across financial documents, checking it against other documents and performing multistep calculations. Its evaluation gives agents PDFs rendered as images. Separate internal tests report gains over K3 in spreadsheet modeling and source-grounded research, where linked formulas and auditable supporting material matter as well as the final answer.

Factual reliability remains material to those uses. Moonshot’s hard-prompt stress test reports a decline in claim-level hallucinations from 23.4% to 8.3%. Artificial Analysis’s separate test reports its own hallucination rate falling from 53% to 44%. Those measures use different tasks and scoring; neither percentage describes the error rate of ordinary financial work.

Lower Token Rates Do Not Fix the Task Bill

Moonshot lists standard K3.1 Pro at $2 per million input tokens, $10 per million output tokens and $0.20 per million cached input tokens. These are US-dollar rates before applicable taxes, each one-third below K3, its predecessor.

The Highspeed variant doubles regular-input and output rates to $4 and $20 per million tokens, while cached input stays at $0.20. Moonshot claims two to three times faster output for that variant. Artificial Analysis lists Sonnet 5.5 at the same regular-input and output rates as standard Kimi, with a lower cached-input rate, so Kimi’s price advantage depends on the alternative and the workload.

Repeated context can reduce input charges when an unchanged request prefix is reused from the cache. Writing new cached context has its own price: $2 per million tokens for five minutes on standard Kimi, or $4 for an hour. A cache hit refreshes the lifetime without another write fee. The bill assigns each input token to one of three categories: a cache hit, a new cache write or regular input.

Consumption can offset the cheaper rates. In its October 23 benchmark report, Artificial Analysis found Kimi’s cost per task roughly unchanged from K3 at $1.98, alongside an increase from about 48,000 to 55,000 output tokens per task. That is a weighted benchmark cost, not an invoice forecast for another application.

Agent-software developer Atomic Agent offers a more concrete, narrower test of effort settings. It ran the same small Python task three times per setup, using five planted bugs, two missing features and 32 hidden checks. Kimi and Opus passed every check in all three runs. At low effort, Kimi averaged 2.2 minutes and $0.16–$0.21; at default maximum effort, it took 4.9 minutes and $0.35–$0.46. Opus averaged 69 seconds and $0.37.

The test reported a discrepancy between Kimi’s account-balance change and token-price arithmetic, hence the cost ranges. On this one launch-day task, extra model calls and reasoning consumed the apparent price advantage at maximum effort. Kimi supports low, high and maximum effort, with reasoning always enabled.

What Users Can Access Before the Weights Arrive

K3.1 Pro and Highspeed are available through Kimi’s hosted products and API. Personal memberships govern access in the apps: the free tier has limited standard-model use, Plus adds the Kimi Code coding tool, and Pro includes Highspeed. Enterprise administrators need to enable the model because organizational access is off by default at launch.

Moonshot promises the full weights by November 4. Those files will let developers run the model on infrastructure they control, subject to the release terms. Kimi’s earlier K3 weights were also released separately from its hosted launch.

The upcoming weights release also carries a security trade-off. Moonshot says it removed cybersecurity-related training material and filtered reinforcement-learning tasks to reduce offensive capability while preserving general coding usefulness. In its internal comparison, an unfiltered model solved 90.5% of its vulnerability tasks within three attempts, versus 81.0% for the final model. The company argues that broadly available models can help more teams audit and repair their own code, including legacy software.

Influencer Allegations About Hidden Sponsorships

On October 23, creator Elshayib posted what he described as Kimi campaign material and said he had declined to use the supplied templates. He said he would not make claims he had not tested and did not want to participate in a China–US political narrative.

In a later October 23 post, he said he was asked to delete a comment after refusing the campaign and criticizing another creator. On October 24, he alleged that participants were choosing existing demonstrations to share, and questioned whether Kimi produced them and whether they were created in one attempt.

Pekingnology’s account describes screenshots requesting draft approval before publication while telling creators to omit paid-partnership labels, website links and official-account tags. The suggested narrative presented Kimi as an open-source counterpart to Opus at a fraction of the price. Pekingnology explicitly says the document has not been independently authenticated; its origin, Moonshot’s authorization and whether anyone published paid endorsements under those terms remain unknown.