Update, October 26, 2026: Kimi K3 now has a successor. Moonshot's announcement dated October 24 introduces K3.1 Pro, a further-trained version of K3, alongside K3.1 Pro Highspeed. Moonshot promises better coding, 3D creation and financial research, and says Highspeed runs two to three times faster than standard Pro. Those are vendor claims. Independently, Artificial Analysis placed K3.1 Pro fourth at launch, scoring 55 on its Intelligence Index.
K3 remains listed in the API model catalog and Kimi Code. The article below explains why K3 mattered when it arrived in July; its original Claude comparison is historical. We have updated the access, pricing, reasoning and weights information where subsequent changes affect readers.
In July 2026, much of the leading AI conversation revolved around ChatGPT, Claude and Gemini. Then Moonshot AI introduced Kimi K3.
Within days, the new model had taken first place on a major front-end coding leaderboard, posted some of the strongest independent agentic-work results available and drawn direct comparisons with Anthropic's then-flagship Claude Fable 5.
At launch, it was offered through a web app, mobile apps, a desktop agent, a coding assistant and an API priced below Fable 5.
That combination raised an obvious question: was Kimi K3 a serious new contender, or another launch being carried by a few flattering benchmarks?
Our launch assessment was that Kimi K3 deserved serious attention. The early evidence did not justify replacing Claude across the board.
The 60-second answer
Kimi is Moonshot AI's assistant and productivity platform. K3 was its flagship model when it launched in July 2026. K3.1 Pro has since taken that position, while K3 remains available through the API and Kimi Code.
K3 was designed for much more than ordinary chat. Its strongest use cases include:
- Long-running coding and software-engineering tasks
- Research across large collections of sources and files
- Creating presentations, spreadsheets, reports and dashboards
- Understanding images and video
- Coordinating multiple AI agents on parts of a larger project
- Processing exceptionally large amounts of information in one session
K3 has 2.8 trillion total parameters, uses a Mixture-of-Experts architecture and supports a one-million-token context window. In practical terms, that gives it the capacity to work across very large codebases, extensive document collections or unusually long projects without immediately losing the thread.
For the newer model, Moonshot lists K3.1 Pro across Kimi's web and mobile apps, Kimi Work, Kimi Code and its API. Access depends on the product and plan. In Kimi Code, the K3.1 Pro model is available on the Plus plan and above. Highspeed and the full one-million-token context require the Pro plan or above. Legacy members have separate tier names and limits. API use is billed separately. Check the model and allowance offered in your account before subscribing.
Kimi is not just another chatbot
The easiest way to understand Kimi is to think of it as an agentic workspace built around Moonshot's models.
The normal Kimi app handles chat, files, images and general questions. Around it, Moonshot has built several more specialized surfaces:
- Kimi Work is a desktop agent that can work with local folders, navigate the web, run code and complete scheduled tasks.
- Kimi Code is a terminal and IDE coding agent capable of reading and editing files, running commands and searching codebases.
- Kimi Slides, Docs and Sheets turn research and instructions into editable work products.
- Kimi Deep Research conducts multi-step research and produces structured reports.
- Agent Swarm can divide larger objectives across multiple specialized agents working in parallel.
Moonshot says Agent Swarm can dynamically coordinate as many as 300 sub-agents and thousands of individual steps. It has also showcased K3 creating interactive research websites, complex presentations, dashboards, software projects and even playable 3D experiences.
Those examples show what Moonshot demonstrated under selected conditions. They do not guarantee a finished professional result from every one-sentence prompt.
Why Kimi K3 drew comparisons with Claude
The strongest launch headline came from Arena's Frontend Code leaderboard, where users compare working web interfaces without initially knowing which model created them.
Kimi K3 reached first place with a score of 1,679, ahead of Claude Fable 5 on 1,631 and GPT-5.6 Sol on 1,618.
That is meaningful because front-end coding and visual web development have been among Claude's most visible strengths. Kimi did not merely perform well on a test designed by its own developer; it won a live, externally operated evaluation based on human preferences.
However, one leaderboard does not make Kimi the world's smartest model.
In our August 13, 2026 review, we recorded Kimi K3 (max) at 60 on the Artificial Analysis Intelligence Index, first among the 103 models displayed then. That is an archived observation, not today's ranking. It should not be compared directly with K3.1 Pro's later score without matching the index version and evaluation settings.
Its agentic results may be more significant. On AA-Briefcase, which evaluates realistic, long-running knowledge-work projects, Kimi K3 ranked second behind Fable 5 and ahead of GPT-5.6 Sol. Its analytical-quality score was effectively tied with Fable, although other models remained stronger in areas such as presentation quality.
Moonshot's Kimi K3 launch announcement acknowledged that its overall performance trailed Fable 5 and GPT-5.6 Sol.
The significance was the competition: another accessible model was getting close to the leaders on work people and businesses paid for.
Kimi K3 vs Claude Fable 5 at launch
This table preserves the July 2026 comparison. Its prices, competitors and rankings describe that launch period, not today's model lineup. The August review recorded above is a separate snapshot.
| Kimi K3 | Claude Fable 5 | |
|---|---|---|
| Context window | 1 million tokens | 1 million tokens |
| API input price | US$3 per 1M tokens | US$10 per 1M tokens |
| API output price | US$15 per 1M tokens | US$50 per 1M tokens |
| Cached input | US$0.30 per 1M tokens | US$1 per 1M tokens |
| Frontend Code Arena | 1st at launch | 2nd at K3 launch |
| Broad independent intelligence at launch | Frontier-level, behind the leading configurations then | Ahead in the launch comparison |
Those launch rates put K3 70% below Fable 5 for standard input, cached input and output tokens.
The comparison was with Claude's flagship at that time, not every Claude tier. The Sonnet 5 introductory offer discussed in our original article was advertised to run through August 31, 2026. That old promotion should not be treated as a current offer.
Current Kimi API prices
As checked on October 26, 2026, Kimi's international API platform lists these rates in US dollars per one million tokens:
| Model | Standard input | Output | Cached input |
|---|---|---|---|
| Kimi K3 | US$3 | US$15 | US$0.30 |
| Kimi K3.1 Pro | US$2 | US$10 | US$0.20 |
| Kimi K3.1 Pro Highspeed | US$4 | US$20 | US$0.20 |
These are usage charges, separate from a Kimi membership, and exclude applicable taxes. Cache writes have their own rates, so the cached-input column is not the full cost of establishing a cache. Lower token rates also do not guarantee a cheaper completed task: a model can use more tokens or require more retries.
For developers still using K2: Kimi says K2-series API access closes to newly registered users on October 31, 2026. Existing users are unaffected by that initial restriction, but retirement of the remaining K2 series is planned for the end of December. This notice does not announce K3's retirement.
Where Kimi K3 looks strongest
1. Coding with a visual component
K3's clearest early strength is work that combines code with visual feedback. It can inspect screenshots, change the underlying code, run the result and continue refining what it sees.
That makes it particularly interesting for websites, interfaces, dashboards, game development and other projects where functional correctness and visual quality matter together.
2. Long-horizon agent work
Many models can solve a contained coding or research prompt. Far fewer remain reliable across hundreds of tool calls and a project that takes hours rather than minutes.
Kimi has been trained specifically for these longer workflows. Its independent knowledge-work results suggest that strength is not purely a marketing claim.
3. Large research and document tasks
The one-million-token context window allows K3 to consider extremely large inputs in a single request. That could include books, reports, legal documents, research papers, lengthy conversation histories or substantial code repositories.
A large context window does not guarantee perfect recall, but it expands the class of work the model can attempt without aggressively summarizing or discarding earlier material.
4. Turning analysis into finished outputs
Kimi is not limited to producing an answer in a chat window. Its workspace can turn research into slides, spreadsheets, documents, interactive widgets and dashboards.
For consultants, researchers, agencies and solo operators, that may be more valuable than a marginal improvement on a reasoning benchmark. The model can potentially carry a task from gathering information to building the thing that gets delivered.
5. Price-sensitive development
For developers running large numbers of agent steps, output-token costs accumulate quickly. K3's lower launch price relative to Fable made it an option worth testing for those workflows.
Automatic context caching also reduces the cost of repeatedly sending the same large codebase or knowledge base to the model.
The weaknesses and catches
Kimi K3's launch has produced plenty of superlatives. Several limitations deserve equal attention.
A launch benchmark win was not an overall victory
K3's strong results in particular tasks did not establish it as the better general-purpose model. At launch, Moonshot acknowledged that its overall performance trailed Fable 5 and GPT-5.6 Sol. That was a comparison with those models at that time, not a verdict on today's K3.1 Pro and Claude lineup.
Artificial Analysis also reported a 51% hallucination rate in its launch evaluation, up from 39% for Kimi K2.6. That result should not be treated as a universal error rate, but it is another reason not to confuse impressive agentic performance with dependable factual accuracy.
It can be verbose and relatively slow
In the launch evaluation cited in our original article, Artificial Analysis measured K3 at approximately 62 output tokens per second and found that it generated around twice the median number of tokens during its intelligence evaluation.
That means Kimi's low per-token price does not always translate directly into a proportionally cheaper completed task. A model that reasons and writes at length can consume more time and more output tokens along the way.
Reasoning effort is now adjustable
K3 launched with maximum reasoning effort. Its current API guide now lists low, high and max settings, with max as the default. Thinking remains enabled. Developers can reduce the effort for lighter tasks, then check the effect on response time and output quality.
It may take too much initiative
Moonshot warns that K3 can make unexpected decisions when instructions are ambiguous because it has been heavily trained for autonomous, long-horizon work.
That matters when an agent can edit files, execute commands or choose the next action without waiting for approval. Users should give it explicit boundaries, check its proposed plan and retain human approval around consequential actions.
Its session history requires care
K3 was trained to preserve its reasoning history across a project. Moonshot says performance can become unstable if an application fails to return that complete history or if a conversation is switched to K3 midway through a session started with another model.
Everyday users inside Kimi's own products may never notice this. Developers connecting the API to third-party agent systems should.
Check the API search route
The K3 API guide still warns against production use of its web-search functionality while it is being updated. Kimi also publishes separate Web Search Basic, Web Search Pro and URL Fetch APIs. These pages do not establish the same readiness for every integration. Check the exact route you intend to use, and verify important sources in research outputs.
K3's weights are available; K3.1 Pro's are promised
Moonshot has published K3's model weights and its Kimi K3 License. Our original July statement that they were still pending is now outdated. Full-model hosting remains a substantial infrastructure undertaking; downloadable weights do not make this a practical home-PC model.
K3.1 Pro has a separate timeline. Moonshot promises its full weights by November 4, 2026. As of this review, that is an announced release, not a completed download or a verified license. K3's license should not be assumed to govern K3.1 Pro.
Anthropic has made serious unresolved allegations
There is also important context behind the Claude comparison.
In February 2026, Anthropic alleged that Moonshot used hundreds of fraudulent accounts to generate more than 3.4 million Claude exchanges in an unauthorized model-distillation campaign. Anthropic said the activity targeted agentic reasoning, tool use, coding, data analysis, computer use and computer vision.
In its September 2026 threat report, Anthropic further alleged that Moonshot forwarded some Kimi customer requests to Claude, presented the responses as its own and retained exchanges for model training. Anthropic said some forwarded requests contained sensitive customer information.
These are Anthropic's allegations, which Choosely has not independently verified. They do not establish that Kimi K3 itself was trained directly on Fable 5 outputs or that K3.1 Pro currently routes requests to Claude. They are relevant context when assessing training provenance and the handling of sensitive work.
What about privacy?
Kimi's consumer privacy policy permits using de-identified input and output to improve its models. It also describes a customer-service request to stop using your content for training. Review the terms for the product and account you actually use; this consumer policy does not establish the terms of a separate API or enterprise agreement.
That does not make Kimi uniquely unsafe, but it does mean users should not assume that a free consumer AI account is an appropriate place for confidential material.
Before uploading client files, internal company data, source code, financial information or sensitive personal records, businesses should review the current policy and determine whether Kimi's enterprise controls satisfy their requirements.
The safest rule is simple: do not give any consumer AI service information you are not authorized or comfortable to share.
Who should try Kimi K3?
K3's launch strengths made it worth testing for people who:
- Build websites, applications or interactive interfaces
- Regularly analyze large reports, research collections or codebases
- Produce client presentations, spreadsheets or consulting-style reports
- Want to experiment with longer autonomous workflows
- Need a capable API without Fable-level pricing
- Are curious about alternatives to the familiar US AI platforms
Our launch assessment was more cautious for quick everyday chat, highly polished creative writing and teams with established production workflows around Claude.
Should Claude users switch?
Not outright.
The sensible move is to test Kimi beside Claude on the work that matters to you. Give both systems the same codebase, research brief or deliverable, then compare:
- Accuracy
- Completeness
- Time to a usable result
- Amount of supervision required
- Quality of the final deliverable
- Total cost rather than advertised token price
Those criteria remain useful when comparing current models. This July comparison cannot establish a winner between K3.1 Pro and the Claude models available today.
The Choosely verdict
K3 mattered because it made Moonshot a credible contender for serious coding, research and work across large collections of documents. Its July results gave users a reason to test another option beside Claude, even when a leaderboard win did not settle the broader comparison.
K3.1 Pro moves that story forward. For someone trying Kimi today, it is the newer model to evaluate, while existing K3 users can test it alongside their working setup. Compare the finished work, time, supervision and total cost before switching. Choosely has not conducted a hands-on K3.1 Pro comparison.
The Change Brief
Get the week’s AI changes in one clear read
Pricing moves, tool launches, free-tier changes and practical stack updates, filtered for people who actually use these tools.
Stay ahead of AI without following it all day. We’ll send you what matters each week.
Continue reading
Related reads
AI Strategy
GPT-5.6 vs Grok 4.5 vs Claude Fable 5: Early Comparison
GPT-5.6, Grok 4.5 and Claude Fable 5 are now available for practical testing. Here is how they compare on pricing, coding, agents and professional work.
AI Strategy
Claude Fable 5 and Mythos 5 Explained: Should Fable Be in Your AI Stack?
Anthropic just released the most powerful AI model it has ever made publicly available. Here is what actually changed, who should care, and the more useful question most of the coverage is...
