Kimi K3 vs DeepSeek V4: which one, and what for
Kimi K3 vs DeepSeek V4: one reads images, the other costs a fraction per token. Every plan carries both, so the choice is price against the work.
Kimi K3 and DeepSeek V4 are both in the LU Labs Cloud chat catalog, and people ask which one to use as if it were a quality question. It mostly is not. The two differ on whether they read images, and on price per token by a factor that decides most real projects.
DeepSeek V4 is two entries rather than one, so this is a three way comparison.
What is in the picker
All three, on every plan and on a credit pack alike. There is no tier in this comparison to get out of the way first: a €5 pack with no subscription puts the same picker in front of you as the top plan does.
Kimi K3 reasons by default, the Think button switches that off, it accepts image input, and it calls tools natively.
DeepSeek V4 Pro 0813 and DeepSeek V4 Flash 0731 both reason with a Think button you control, both call tools natively, and neither accepts image input.
Price per million tokens
Credits are the unit you actually hold, and one credit is $0.00001. The catalog price converts straight into them, so here is what a million tokens draws on each.
| Model | Input, per 1M | Output, per 1M |
|---|---|---|
| Kimi K3 | 285,000 credits | 1,425,000 credits |
| DeepSeek V4 Pro 0813 | 130,000 credits | 260,000 credits |
| DeepSeek V4 Flash 0731 | 8,000 credits | 18,000 credits |
Read the output column, because output is the side that moves. K3 costs about five and a half times what V4 Pro costs per output token, and about eighty times what V4 Flash costs. That is not a rounding difference you can ignore in an agent loop.
For scale: a €5 pack of 165,000 credits is roughly 115,000 output tokens on K3. The same pack is over nine million output tokens on V4 Flash, which it does reach, and that spread is the whole argument of this page.
Thinking, and who controls it
All three reason, and on all three the Think button is a real switch rather than a label. That matters more here than the raw prices suggest, because a reasoning pass you did not need is billed at the output rate above.
The pattern that holds across the three: leave thinking on for work that has to be derived, and off for extraction, reformatting, translation and summarising. On K3 that single habit is worth more than any model choice on this page.
Vision
K3 reads images. Neither DeepSeek V4 entry does; the upstream refuses image input on both.
If your loop involves screenshots, that decides it, and no price advantage recovers a capability the model does not have. If your work is text in and text out, this line does not apply to you at all and you should be reading the price table instead.
Context
Both DeepSeek V4 entries carry a one million token context, and so does K3. This is the one axis where the three do not separate, which is worth saying plainly because it is the spec people compare first.
How to actually choose
If the work is a long loop, use V4 Flash. At 18,000 credits per million output tokens it is the cheapest way to generate the intermediate output an agent produces and nobody reads. Move the final pass to something larger if the output quality decides something.
If the work is one careful pass over a long document, V4 Pro sits in the middle: five times cheaper than K3 per output token, with the same context window and the same controllable reasoning.
Reach for K3 when the problem is genuinely hard, and use the Think button when you do. It is a capable model and it is also the most expensive thing in the catalog per output token, so the habit to avoid is running long agent sessions on it by default.
And if any part of the job involves an image, that is K3, whatever you are paying.
The part that is not in the table
Credits are one pool for chat, images, video and audio together, so a model comparison is only ever half the budget question. A single image render costs more than a long chat turn, and a video clip costs more again. The pricing page has the rates side by side.
If you want the step by step on getting K3 running, including the API key and the endpoint, that is in how to use Kimi K3 online. The short version, with the plan and the price, is on Kimi K3 online.