Is Kimi AI actually good? This review skips the marketing and looks at what happened when independent testers gave Kimi real work: coding jobs, research, long files and tricky edge cases. Every result names its source.
Our reading of the published results. The labels are a summary, not a score.
Kimi is a very capable assistant that does its best work on long jobs. Its top model, Kimi K3, ranks among the leading AI models in independent testing and is especially good at everyday coding and front-end design. The trade-offs are real: it is slower than most rivals, it stumbles more often on tricky tasks, and it is not cheap for developers. For most people the free plan is a no-risk way to find out if it suits them.
These are results other people measured and published. Most are for Kimi K3, the model on paid plans. The colour shows how Kimi did: green is strong, blue is good, amber is mixed, pink is weak.
0.2 points behind out of 70
An independent engineering channel had Kimi K3 fix real GitHub issues: plan, write the code, then check it. A leading closed model, Claude Opus 4.8, scored 64.3 out of 70. Kimi K3 was about 0.2 points lower, which is close to a tie.
Source: MindStudio write-up of an independent test, 25 July 2026
Just above 60 out of 70
On larger builds in the same test, the closed model scored 62.2 and Kimi K3 scored just over 60. The gap was real but moderate, and Kimi cost less per task.
Source: MindStudio write-up of an independent test, 25 July 2026
36% failed vs about 8%
The tester also wrote "trap" tasks with hidden rules and false starting points. Kimi K3 failed 36% of them. The closed model failed about 8%. The traps were designed around Kimi's known weak spots, so everyday work should go better than this.
Source: MindStudio and Geeky Gadgets, reporting the same test, July 2026
Ranked first
On a public leaderboard where people vote on web pages built by different AI models without knowing which is which, Kimi K3 was ranked first for front-end code after launch.
Source: Arena front-end ranking, as reported in a 20 July 2026 review roundup
3rd of 117
Artificial Analysis runs the same set of hard tests on every model. Kimi K3 placed third of 117 models in its class on their Intelligence Index.
Source: Artificial Analysis, Kimi K3 page, checked 8 October 2026
41.7 tokens a second median 70.8
Artificial Analysis measured Kimi K3 as "notably slow". It also waits longer before it starts: about 3.5 seconds to the first word, against a typical 2.3 seconds.
Source: Artificial Analysis, Kimi K3 page, checked 8 October 2026
10 to 25 minutes per report
Kimi says a typical Deep Research report takes 10 to 25 minutes and cites about 26 sources. Early users in one review roundup said research tasks were done well but noticeably slower than rivals.
Source: Kimi help centre for the figures. Review roundup, 20 July 2026, for the user reports
Check its work
Reviewers reported that Kimi K3 sometimes states wrong things with confidence, and that its results are uneven from task to task. This is true of all AI models to some degree.
Source: Review roundup, 20 July 2026
Test results age quickly. Models are updated often, and a score from July may not match what you see today. Sources are listed in full under Sources and method.
When the job is well described, Kimi performs close to the most expensive models. Say exactly what you want and what "done" looks like.
Hidden rules, unusual cases and wrong assumptions are where it slips. Review its work, run the tests, and ask it to explain its steps.
Kimi is built for depth, not speed. It suits jobs you can start and come back to. For quick answers, the free Kimi K2.6 chat is the faster choice.
Kimi is more than a chat box. We could not find independent tests of each tool, so this table describes what each one does and what to watch for, based on Kimi's own help pages.
| Tool | What it does well | What to watch |
|---|---|---|
| Deep Research | Long, organized reports with sources. Asks what you mean before it starts | Takes 10 to 25 minutes. One report uses about 5 to 10% of a month's credits on the $19 plan |
| Website builder | Builds and publishes a working site from a description. Every version is saved | Uses a lot of credits. Address ends in .ok.kimi.link unless you host the code yourself |
| Kimi Code | Works inside your project. Asks before it changes files or runs commands | Needs a paid plan. Has usage limits that reset over time |
| Slides and Sheets | Turn a prompt or a research report into a finished deck or spreadsheet | Check numbers and formulas before you share them |
| File reading | Up to 50 files of 100 MB each. Kimi K3 reads about a million tokens at once | Free chat uses Kimi K2.6, which holds much less. Scanned PDFs are harder |
| Apps | Web, Windows, Mac, Android and iPhone | The Mac app needs an Apple chip. Desktop Work mode is still in beta |
On a small screen, swipe the table sideways to see every column.
The free plan gives you everyday chat with Kimi K2.6 and a few agent credits. The $19 Moderato plan adds Kimi K3, Deep Research, the website builder and Kimi Code. That is in line with rival assistants at about $20 a month.
The catch is credits. Heavy tools drain the cheaper plans quickly, so the $19 plan suits weekly use, not daily heavy use.
Kimi K3 costs $3 per million tokens in and $15 out. Artificial Analysis calls that expensive for an open model, and it also notes Kimi K3 writes long answers, which adds cost.
Against the most expensive closed models it can still work out cheaper per finished task, as the coding test above found. Kimi K2.6, at $0.95 in and $4 out, is the budget choice for routine work.
Comparing options? See Kimi vs ChatGPT, Kimi vs DeepSeek and Kimi vs Perplexity.
The best review is your own. These five tasks match the kinds of work tested above, and all of them run on the free plan.
Write a Python function that takes a list of email addresses and returns only the valid ones, with duplicates removed and the original order kept. Include three test cases.
Build a single HTML file with a tip calculator: bill amount, tip percentage slider and number of people. Show the tip and the amount per person. Make it look clean on a phone.
Compare heat pumps and gas boilers for a small home: cost to install, cost to run and typical lifespan. Put it in a table and say where each number comes from.
A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? Explain your steps, then check your own answer.
Attach a long PDF you know well, then send: Summarize this in 6 bullet points with the page number for each, and list anything you were unsure about.
Tick what Kimi got right. Four or five ticks means it is likely a good fit for your work. Your ticks are not saved.
We read published tests and measurements of Kimi, kept the ones that described real tasks and gave figures, and reported them with their limits. Where sources disagreed, we left the number out.