QUICK ANSWER:
Kimi is Moonshot AI’s open-weight flagship: a 2.8-trillion-parameter mixture-of-experts model with native vision and a one-million-token context window. Its commercial hook is flat pricing: one rate across the whole window with no long-context surcharge, and cached input at about a tenth of the standard rate. That makes whole-document-set analysis practical, and because output costs roughly five times input, it rewards being used as a reading model rather than a writing one.
Long context windows have been oversold for two years. Vendors announce a million tokens, everybody writes an article about how you can now feed it a whole book, and then the pricing usually makes it impractical.
Kimi is where that stopped being true, and the reason is a pricing decision rather than a technical one.
What Kimi is
Kimi charges the same rate for a token at position 900,000 as it does at position 900. Most providers add a long-context premium, so the model that reads your entire archive is usually the expensive one. Here it is not.
Moonshot AI publishes Kimi as an open-weight flagship: a mixture-of-experts model measured in trillions of parameters, with native vision input, a 1,048,576-token context window and reasoning enabled by default rather than as an optional mode. The weights go up on the usual hubs at each release, and it has repeatedly set the high-water mark for the largest openly released model.
“Open weights” here means downloadable, not trivially runnable. At this scale, self-hosting is a data-centre exercise. The practical route for a business is the API, and for most people the open weights matter as insurance rather than as a deployment plan: the model cannot be withdrawn from under you, and someone else can always host it.
The pricing, and why flat matters
Three numbers describe Kimi’s pricing, and none of them is a dollar figure. Cached input costs about a tenth of uncached input. Output costs about five times input. And every one of those rates holds flat across the full context window, with no tiering once a request gets long. Headline rates move between releases; that structure has not.
Compare that with the usual industry pattern, where the rate steps up once a request passes a threshold, often around 128,000 tokens. Under that model, the request you most want to make, the one containing everything, is the one that costs disproportionately more. Flat pricing removes the penalty, and it changes which jobs are worth doing.
A worked example. Say you want to analyse a full set of supplier contracts, about 700,000 tokens of text, and produce a two-thousand-word risk summary. That is 0.7 million input tokens and roughly 3,000 output tokens, in a single call at the standard flat rate. Priced against a person reading the same pile, it rounds to nothing. The alternative used to be chunking the documents into twenty separate calls and stitching the answers together, losing the cross-document context that was the whole point.
The ratios also tell you where the cost sits. Output runs about five times input, and because reasoning is always on, it generates thinking tokens you also pay for. That makes it a superb reading and reasoning model. Pair it with a cheaper model when you need volume writing, and you get the best of both.
Where it is strong
Kimi consistently places near the top on software engineering and coding benchmarks, and has led several long-horizon coding evaluations outright. Individual scores age badly and are not worth quoting, but the pattern has held across a broad set of them for long enough to trust the general claim: this is a strong coding and analysis model.
Three business applications follow naturally from the combination of long context and analytical strength.
Whole-codebase review
A million tokens holds a meaningful proportion of a small application. Asking a model to review a whole codebase rather than a file at a time produces genuinely different answers, because most of what matters lives in the relationships between files. If you maintain a custom system, this is the highest-value use of the model, and it pairs well with the kind of technical review that goes into a web build.
Document set analysis
Tenders, leases, insurance policies, years of email on one matter. The value is not summarising each document, which any model can do, but answering questions that require all of them at once, such as which clauses are inconsistent between agreements.
Research synthesis with vision
Native vision means charts, scanned pages and screenshots go into the same request as the text. For anything where the important detail is in a figure rather than a paragraph, that removes a preprocessing step that usually loses information.
The one setting that cuts your bill
Flat pricing makes one optimisation easy to overlook: you do not have to resend the whole context every time. If you are running a conversation or an agent loop over the same large document set, you are paying the full input rate for the same tokens repeatedly.
Cached input exists precisely for this, and at roughly a tenth of the uncached rate it is the largest single lever on your bill. Structure requests so the stable material sits at the front, unchanged between calls, and the variable part goes at the end. Get that right and a long-running analysis costs a tenth of the naive version. This is the same optimisation that dominates agent costs, as covered in our Hermes Agent guide.
Kimi compared
| Model | Relative input cost | Output vs input | Context | Best for |
|---|---|---|---|---|
| Kimi | Highest of the three | About 5x | 1M, flat rate | Reading and reasoning over very large inputs |
| Qwen Max-class | About two thirds of Kimi | About 3x | 1M, flat rate | General work where output volume is high |
| DeepSeek Pro | Lowest by a wide margin | About 3x | 1M | Cheapest capable reasoning at volume |
Australian considerations
Moonshot AI is a Chinese company and the hosted API processes data offshore, which makes it a cross-border disclosure under Australian Privacy Principle 8 with the accountability that carries. The document-analysis use cases in this article are exactly the ones most likely to involve personal or commercially sensitive information, so decide the boundary before you build the workflow rather than after.
The mitigation that works in practice is redaction upstream. Run a local model to strip names, addresses and identifiers, send the de-identified text to Kimi for analysis, and map the results back locally. It adds one step and keeps the sensitive material entirely in your hands.
One more scheduling note: Moonshot has retired earlier generations of its API on published sunset dates. If you have anything built against the older endpoints, that migration is due.
Pros and Cons
Pros
- Flat pricing across the full one-million-token window, with no long-context premium.
- Cached input at one tenth of the standard rate.
- Top-tier coding and software engineering performance.
- Native vision, so charts and scans go in with the text.
- Open weights, so the model cannot simply disappear.
Worth knowing before you start
- Output is priced well above input, so it suits reading more than writing.
- Reasoning is always on, so thinking tokens count toward your bill.
- Large enough that self-hosting is a data-centre exercise, not a desktop one.
- Offshore processing, with APP 8 obligations on the Australian business.
- Older API generations get retired on a schedule, so check your integrations.
The verdict
Kimi is the model to reach for when the question requires reading a great deal before answering, and when getting the answer right is worth more than the price of the call. Contract review, codebase analysis, multi-document investigation. For those it is excellent value, because the alternative is a person doing it.
For writing volume and routine drafting, hand the work to a cheaper model and keep Kimi for the thinking. Used that way it is one of the best value calls you can make.
Put it to work on something that matters
The businesses that get the most from long-context models are the ones that point them at a genuinely time-consuming document job first, rather than experimenting on something trivial. If you want help identifying which of your processes is worth pointing this at, our AI consulting team runs an AI Jumpstart Audit that does exactly that, and our AI training workshops teach teams to run the analysis themselves.
One real document set through the API will tell you more than any benchmark. You can find it at Kimi.
Frequently asked questions
How is Kimi pricing structured?
Output costs about five times input, cached input costs about a tenth of uncached input, and every rate is flat across the full one-million-token context window with no surcharge for long requests. That flat structure is the reason whole-archive analysis is practical here. Moonshot lists the current per-token figures in its API docs.
What is the Kimi context window?
1,048,576 tokens, which is roughly 750,000 words of English. In practice that is a full contract set, a small codebase or several years of correspondence on one matter, all in a single request.
Is Kimi open source?
The weights are openly published, and it has repeatedly set the high-water mark for the largest openly released model. It is not practical to self-host at 2.8 trillion parameters outside a data centre, so most businesses use the hosted API and treat the open weights as protection against the model being withdrawn.
What is Kimi best at?
Reading and reasoning over very large inputs, and software engineering tasks. It scores 67.3 on SWE-bench under the mini-SWE-agent harness and ranks near the top across several coding benchmarks. For high-volume writing, pair it with a cheaper model, because output costs around five times what input does.
How do I keep Kimi costs down?
Use the cache. Put the stable material at the start of the request so it stays unchanged between calls, and keep the variable part at the end. Cached input is one tenth the price, which on a long-running analysis is the difference between a sensible bill and a startling one.
