Moonshot's Kimi K3.1 Pro flirts with open weights leadership amid paid influencer scandal allegations

Chinese AI startup Moonshot just unveiled its latest and most large language model (LLM) yet on Friday: Kimi K3.1 Pro is a 3-trillion parameter post-trained version of its predecessor Kimi K3 released back in July, and it looks to be a historic release.

That's because Kimi K3.1 Pro scores so high on third-party benchmarks — according to both Moonshot and third-party AI model testing firm Artificial Analysis — it will almost certainly take the crown as the top open weights AI model in the world when Moonshot releases the weights under an open license, which the company says will happen a week and a half from now, on November 4, 2026.

Already, Kimi K3.1 Pro dethrones Xiaomi’s MiMo-V2.6-Pro, DeepSeek V4.1-Flash-1017, Meta's new Muse Wonder, and even OpenAI's GPT-6 Astra on Artificial Analysis's Intelligence Index.

The model excels at complex coding, financial analysis, 3D creation, and software-security work, showing particularly large improvements in tasks requiring repeated coding, tool use, debugging, and revision.

For now, developers and users can access it through its applications and developer application programming interface (API), and its rates are half those of Claude Opus 5.5 and Meta’s Muse Wonder for input and output, although DeepSeek and Xiaomi remain cheaper.

However, it has also become the subject of an alleged paid influencer scandal on the social network X, as well.

The biggest gains are in sustained coding and tool-driven work

Kimi’s most dramatic improvements appear in evaluations where an agent must execute a sequence of actions, rather than simply answer a difficult question.

The general results come from Artificial Analysis; its Cyber Index and launch results supply the security comparisons. Percentages, index points, and Elo ratings measure different things and should not be compared across rows.

For development teams, the terminal result makes Kimi worth evaluating on repository maintenance, debugging, and workflows that alternate between inspecting files, editing code, running tests, and responding to failures. The business-software result supports testing agents that perform several connected operations. These are implications for pilot selection, not measured guarantees for a company’s own applications.

The evaluator’s current Kimi–Opus comparison also puts Kimi ahead on AutomationBench-AA, 73% to 70%, and GDP.pdf document reasoning, 31% to 26%. Opus retains leads on Humanity’s Last Exam, 61% to 52%, and CritPt, 32% to 25%. That split favors choosing by workload rather than treating the aggregate ranking as a universal ordering.

The gains are smaller on other independent tests: Humanity’s Last Exam rises from 47% to 52%, the CritPt physics evaluation from 23% to 25%, and SciCode from 59% to 61%. The evidence points more strongly to improved task execution than to an equally large leap in every kind of reasoning.

Moonshot’s own comparison provides another perspective:

Moonshot assembles these comparisons from its own testing and externally reported scores, with differences in agent software and evaluation conditions. They are not a single controlled head-to-head trial.

The company’s Terminal-Bench result also differs from Artificial Analysis’s: Moonshot puts Kimi narrowly behind Opus, while the independent evaluator reports a leading 67%. Both results support competitiveness; neither establishes universal coding superiority.

For scientific-computing teams, the Science result is a reason to retain alternatives in the evaluation. Kimi trails Astra by 30 points and Opus by 25.2 points in Moonshot’s comparison, despite performing strongly on other coding measures.

API pricing: cheaper than Opus, priced alongside Sonnet, above DeepSeek

Moonshot’s API platform lists K3.1 Pro at $2 per million input tokens, $10 per million output tokens, and $0.20 per million cached-input tokens. Those rates are one-third below Kimi K3’s.

The Highspeed variant costs $4 for input and $20 for output, with cached input still $0.20. Moonshot claims equivalent intelligence at roughly two to three times the speed. Faster model generation should be evaluated separately from the duration of a full workflow that also waits for tools, builds, and tests.

These are published USD token rates checked October 24, 2026, excluding additional tool fees, cache-write charges, taxes, and negotiated discounts. The listed OpenAI rates apply through 272,000 input tokens; longer requests carry higher rates. DeepSeek’s documentation maps its Flash API to the 1017 model and applies off-peak prices outside designated weekday hours, including all weekend hours. GPT-6 Sol is included for the cyber comparison; its documentation identifies GPT-6.1 Sol as a newer model.

Standard Kimi costs 50% less per input or output token than Opus 5.5 and Muse Wonder, and 80% less than Astra’s listed base rates. But it offers no input/output discount against Sonnet 5.5, which scores one point higher on the broad Intelligence Index. Sonnet also charges less for cached input.

Against DeepSeek at peak rates, Kimi charges about 6.7 times as much for uncached input and 8.3 times as much for output. The case for paying that premium is better results on a particular workload, not a general claim that open-weight-oriented models are inexpensive.

Unreal Engine, CAD, and financial models: what Moonshot says it can build

Moonshot’s coding showcase explicitly identifies Unreal Engine alongside Unity and Godot. The company says Kimi can combine game-engine work with web research and 3D generation to produce complete interactive experiences. It also highlights browser interfaces, Three.js scenes, and improvements in layout, animation, and visual presentation.

For businesses, those claims suggest pilots involving interactive product demonstrations, training simulations, game prototypes, and browser-based visualization. They do not establish that an API request alone delivers a production-ready Unreal project: the surrounding agent still needs access to the engine, project files, tools, and a way to test the output.

The 3D work extends beyond games. Moonshot says Kimi can turn a written brief, sketch, or reference image into a model; work with existing asset collections; and modify geometry, materials, and scene arrangements through Blender or CAD scripts without requiring MCP connectors for those scripted operations.

It specifically highlights adjustable designs for mechanical components, architecture, configurable products, and furniture. Its 3D Make application lets users inspect and edit models online, download them for 3D printing, or export videos and Live Photos.

That makes design exploration and prototype generation plausible applications. The BenchCAD result measures reconstruction quality, however, not manufacturing tolerances, structural integrity, or regulatory approval. Engineering teams would still need their existing validation processes before treating a generated part as ready to manufacture.

The distinction is consequential: the reported agent investigated defects in the environment instead of endlessly tuning the behavior it rewarded. This remains a company demonstration, not independent evidence of comparable reliability across engineering projects.

Finance is another explicit target. Moonshot describes spreadsheet modeling, investment screening, profitability analysis, and research that traces conclusions back to evidence. Its internal tests report gains of 7.2 points in financial modeling and 11.6 points in research, plus approximately 4.2 times its predecessor’s score in a simulated business-management exercise.

The independent GDPval-AA result and Moonshot’s OfficeQA result provide reasons to test document-heavy analyst workflows. Appropriate pilots could include preparing a sourced first draft, checking figures across filings, and assembling a workbook for an analyst to review. The internal simulated-management result does not establish a corresponding increase in real-world revenue or productivity.

More post-training — and an agent that finds faults in its simulator

Moonshot attributes the upgrade to substantially more post-training of Kimi K3, including reinforcement learning across hundreds of task domains. This is the stage in which models learn from feedback on attempted tasks.

The company describes improvements throughout that process: the training software, long-context processing, computing efficiency, isolated tool environments, scheduling of attempts, and checks on environment quality. Its explanation centers on making larger-scale training work consistently; it does not identify a newly invented model architecture or quantify a training-cost breakthrough.

A reported walking-simulator experiment gives a concrete example of the behavior Moonshot seeks.

Kimi built a simulation and training pipeline around human-atlas, an anatomical model with 296 skeletal elements and 402 muscle components. Across 167 missions and 420 independent reviews, Moonshot says it revised the controller until the simulated body progressed from falling within seconds to alternating steps for 11.2 seconds.

When reward adjustments failed to improve arm movement, the model investigated the simulator’s dynamics, corrected 36 shoulder-muscle moment-arm calculation errors, and retrained the controller.

The enterprise relevance is the debugging pattern: investigate whether the environment is wrong, rather than repeatedly adjust the same unsuccessful solution. That is useful behavior to look for in engineering agents. This particular result remains a company demonstration, rather than proof of repeatable performance on unfamiliar industrial simulations.

Cyber defense is a strength, but the aggregate score hides trade-offs

The Cyber Index evaluates finding vulnerabilities, reproducing them, and making repairs without breaking existing software. Kimi’s 64% aggregate result gives security teams a reason to test it for code-review and remediation assistance, including difficult repositories that require repeated investigation.

Its component scores show why model selection still depends on the task.

Kimi leads CyberGym-E2E-AA at 95%, compared with Daybreak Blue’s 93%. On CWE-Bench-AA, DeepSeek leads at 69%, ahead of Daybreak Blue at 65% and Kimi at 61%. On DeepsecBench-AA, Kimi scores 37%, behind Daybreak Blue’s 48%.

Artificial Analysis also reports no safety refusals from Kimi on these tasks. Some competitors lose points when their services decline requests. The ranking therefore reflects both capability and what the tested service permits developers to do; it does not isolate a model’s maximum technical ability.

Moonshot says it separately restricted offensive-security capabilities by removing cybersecurity-related training material and filtering reinforcement-learning tasks. On its internal exploit-development evaluation, the unfiltered version solved 90.5% of problems with up to three attempts, compared with 81% for the final model. The company says broader coding improvements still transfer to security work despite those interventions.

For deployment, these results support a supervised vulnerability-research or patch-proposal pilot. They do not establish that an agent can safely apply every proposed repair to live infrastructure without tests and review.

Cost per completed workflow matters more than the token discount

Artificial Analysis’s launch report puts Kimi at $1.98 per Intelligence Index task, approximately unchanged from K3. Average output rises from roughly 48,000 to 55,000 tokens, helping absorb the lower rates.

Its live comparison page, checked October 24, instead lists $2.11 for Kimi and $5.98 for Opus 5.5. The launch and live Kimi figures differ, but the latter comparison still puts Kimi’s measured cost about 65% below Opus’s; it is not a forecast of savings on every enterprise task.

The cyber evaluation shows a more pronounced trade-off:

Artificial Analysis attributes Kimi’s approximately 15% cost increase over K3 to roughly doubled input-token consumption. Against DeepSeek, it gains seven percentage points while costing approximately 6.7 times as much per evaluated task.

That suggests a concrete deployment experiment: use a less expensive model for work it handles satisfactorily and test Kimi on unresolved or more difficult cases. Whether that routing saves money depends on the cost of retries and failures; the published results do not measure that proposed arrangement.

The Kimi billing documentation also says repeated prefixes use automatic caching, normally with a five-minute lifetime, and that cache writes have separate charges determined by retention duration. Teams replaying large repositories or documents should measure cache writes and hits, not estimate the bill from output alone.

For an enterprise pilot, the useful comparison is cost per accepted deliverable, including tool execution, retries, and human correction, alongside completion time and failure rate.

Developers need to check more than the model name

The API identifiers are kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Moonshot’s parameter reference says both follow Kimi K3’s constraints.

Reasoning effort accepts low, high, or max, defaulting to max. Changing effort during a conversation invalidates prefix-cache hits. Developers should benchmark the intended setting rather than assume lower-effort calls reproduce maximum-effort launch scores.

The documentation also fixes temperature at 1.0 and top-p at 0.95, and supports automatic, disabled, or required tool selection. For multistep exchanges, applications must preserve complete returned assistant messages, including reasoning content.

These details matter when adapting an existing agent: retained conversation state, tool handling, and cache behavior can affect both results and cost. An API model also needs the surrounding execution environment to reproduce the company’s code, CAD, and game-engine demonstrations.

Moonshot lists access through its web and mobile applications, Work, and Code, and says API customers can use Zero Data Retention. Enterprise administrators must enable K3.1 Pro because organizational access starts disabled.

Allegations of undisclosed influencer promotion complicate the launch

Kimi K3.1 Pro’s debut has also drawn scrutiny over allegations that influencers were asked to promote the model without disclosing a commercial relationship to Moonshot.

X user Elshayib posted screenshots he described as instructions for a Kimi promotional campaign, saying he declined to participate on those terms. The purported instructions required creators to submit drafts for approval and included the direction: “Please DO NOT put on the paid partnership label,” which would be a violation of various U.S. laws and likely those of other companies.

Suggested talking points portrayed Kimi as an open-source equivalent of Anthropic’s Claude Opus 5.5 at a fraction of the price.

Crucial facts remain unresolved. As Pekingnology’s reporting explicitly notes, the document has not been independently authenticated. It is unclear whether the instructions originated with Moonshot, an outside agency or another party; how many creators received them; or whether anyone published compensated endorsements under those terms.

Circulation of the screenshots does not establish that Moonshot authorized an undisclosed advertising campaign. No confirmed company response to the specific allegations was identified in the material reviewed for this subsection as of October 25.

The disclosure issue nevertheless matters. X’s policy requires its paid-partnership disclosure on qualifying posts. U.S. Federal Trade Commission guidance also calls for clear disclosure of material connections between endorsers and brands, including payments or free products and services. Whether a particular endorsement violated those standards would depend on the actual relationship, content and disclosures—not simply the existence of a purported campaign brief.

For enterprise buyers, the practical concern is whether apparently independent recommendations reflect undisclosed incentives. Procurement teams should ask reviewers about payment, credits and editorial approval, and verify pricing and deployment claims against published terms. With K3.1 Pro’s weights still pending and its production API carrying usage charges, sweeping descriptions such as “open source” or “free” need qualification.

The allegations do not, by themselves, invalidate independent benchmark results. They do reinforce the importance of separating promotional enthusiasm from reproducible evidence: performance on a company’s own workloads, total operating costs, data-handling terms and the availability of deployable weights.

Where Kimi belongs on an enterprise shortlist — and where it does not yet fit

The launch evidence supports several targeted evaluations, rather than a blanket replacement of existing models.

This is an assessment of where to pilot the model, based on the published evidence; the launch does not provide production acceptance rates for these enterprise workflows.

Scientific workflows deserve particular caution given the large gap in Moonshot’s Science comparison. High-volume routine tasks also need a strong quality justification for paying Kimi rates when DeepSeek and MiMo are substantially cheaper.

Nor do the results justify unreviewed financial conclusions or automatic production changes. Artificial Analysis reports Kimi’s AA-Omniscience hallucination rate falling from 53% to 44%, while accuracy rises from 48% to 51%. Moonshot’s separate internal stress test reports a decline from 23.4% to 8.3% at the claim level. Those are different measurements, and neither is an everyday error-rate guarantee.

Finally, organizations requiring immediate self-hosting cannot meet that requirement with K3.1 Pro yet. The final weights, license conditions, and production deployment requirements remain outstanding. Zero Data Retention for a hosted API does not establish that a particular data-location, contractual, or infrastructure requirement is satisfied.

Until the weights arrive, enterprises using Moonshot’s API must entrust their inputs to an external service whose international privacy policy identifies Singapore-based servers, rather than assume processing stays in the United States. That may give government suppliers particular pause amid tighter scrutiny of Chinese-developed AI: following restrictions covered by VentureBeat, a DHS order excludes DeepSeek from civilian federal procurement and requires removal of its products and covered third-party implementations from agency systems. That order does not target Kimi, but gives prospective users another reason to verify processing locations and government-contract requirements before sending sensitive material.

The strongest reason to consider Kimi is its combination of competitive multistep performance, lower standard token prices than several flagship rivals, and a promised route to operating the model independently. The strongest reason to keep alternatives is equally concrete: cheaper models may already satisfy the task, proprietary rivals remain stronger on some evaluations, and self-hosting is still a future option.

For enterprise teams, the next step is a bounded comparison on their hardest recurring work—measuring accepted results, total cost, and time saved—while waiting for November’s weights release to determine whether the deployment flexibility matches the benchmark promise.