Model comparison
Kimi K3 vs DeepSeek V4 Pro: pricing, context and features compared
Kimi K3 costs $0.80 per million input tokens and $10 per million output tokens; DeepSeek V4 Pro costs $0.2088 and $0.4176. Here is how they compare on specs, on what real workloads cost, and on which to pick for what, from prices updated daily.
Kimi and DeepSeek’s Pro model are both large open-weight models with very long context windows. They differ in what they accept and how they price output.
- Input / 1M
- $0.80
- Output / 1M
- $10
- Context
- 1,048,576
- Max output
- Not published
Open weights: this is the price of OpenRouter’s top-ranked host, which may differ from Moonshot AI’s own API price.
Full Kimi K3 pricing and specs- Input / 1M
- $0.2088
- Output / 1M
- $0.4176
- Context
- 1,024,000
- Max output
- 384,000
This page covers the original DeepSeek V4 Pro weights (0423), and its price is the top-ranked host’s on OpenRouter, not DeepSeek’s. DeepSeek’s own API now serves the newer V4-Pro-0813 under the deepseek-v4-pro name, at $1.32 input and $3.96 output per million tokens during peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays) and half that, $0.66 and $1.98, at all other times.
Full DeepSeek V4 Pro pricing and specsVerdict
Kimi K3 or DeepSeek V4 Pro: which to pick
DeepSeek V4 Pro has the lower list prices: 3.8× cheaper per input token and 24× per output token. On our data, Kimi K3 has the edge for non-text inputs, and DeepSeek V4 Pro for the lowest bill, very long prompts and repeated prompts. For scale, a support chatbot handling 1,000 requests a day costs about $152.46 a month on Kimi K3 and $10.24 on DeepSeek V4 Pro.
- Lowest bill for typical workloadsPick DeepSeek V4 Pro
- DeepSeek V4 Pro costs less on all 4 workloads below, 6.5× to 15× cheaper than Kimi K3 at list prices.
- Very long prompts (whole codebases, long documents)Pick DeepSeek V4 Pro
- Both read about 1,024,000 tokens per request. One 810,000-token prompt with a 2,000-token reply costs $0.67 on Kimi K3 and $0.17 on DeepSeek V4 Pro. Neither has a long-prompt surcharge in our data.
- Long replies (reports, large code files)Either
- DeepSeek V4 Pro caps a reply at 384,000 tokens; Kimi K3 publishes no separate cap, so we can’t say which allows longer replies.
- Images, audio, video or files in the promptPick Kimi K3
- Kimi K3 also accepts images and video, which DeepSeek V4 Pro doesn’t.
- Repeated long prompts (prompt caching)Pick DeepSeek V4 Pro
- DeepSeek V4 Pro bills cached input at $0.0174 per 1M (8% of its input price), against $0.55 per 1M (69% of its input price) for Kimi K3.
- Overnight batch jobsEither
- Neither has a batch price in our data, so jobs that can wait cost the same as real-time requests.
- Self-hosting or choosing your own hostEither
- Both have open weights, so you can run either yourself or buy it from several hosts; prices here are typical hosted prices.
Criteria use only list prices, published limits and the features each provider declares. We don’t rank answer quality; test both models on your own prompts before you commit.
Specs
Kimi K3 vs DeepSeek V4 Pro side by side
| Spec | Kimi K3 | DeepSeek V4 Pro |
|---|---|---|
| Provider | Moonshot AI | DeepSeek |
| Input / 1M tokens | $0.80 | $0.2088 (better) |
| Output / 1M tokens | $10 | $0.4176 (better) |
| Cached input / 1M | $0.55 | $0.0174 (better) |
| Batch in / out | No batch price | No batch price |
| Long-prompt price | None | None |
| Context window | 1,048,576 | 1,024,000 |
| Max output | Not published | 384,000 |
| Accepts | text, images, video | text |
| Image input | Yes (better) | No |
| Tool calling | Yes | Yes |
| Reasoning mode | Yes | Yes |
| Prompt caching | Yes | Yes |
| Structured output | Yes | Yes |
| Open weights | Yes | Yes |
| Knowledge cutoff | Not published | Not published |
| First listed | 2026-07-16 | 2026-04-24 |
Prices in US dollars per million tokens. Highlighted: the lower price or larger limit where the difference is over 5%. Features are as declared by each provider’s API; “First listed” is when our data source first listed the model. Long-prompt prices apply to the whole request once the prompt passes the threshold.
Costs
What Kimi K3 and DeepSeek V4 Pro cost for real workloads
| Workload | Tokens in / out | Requests/day | Kimi K3 / month | DeepSeek V4 Pro / month | Cheaper |
|---|---|---|---|---|---|
| Support chatbotCalculator: Kimi K3 · DeepSeek V4 Pro | 1,500 / 400 | 1,000 | $152.4650% cached | $10.2450% cached | DeepSeek V4 Pro (15×) |
| RAG appCalculator: Kimi K3 · DeepSeek V4 Pro | 6,000 / 500 | 500 | $144.4820% cached | $18.7420% cached | DeepSeek V4 Pro (7.7×) |
| Coding agentCalculator: Kimi K3 · DeepSeek V4 Pro | 40,000 / 2,000 | 200 | $267.6780% cached | $18.6380% cached | DeepSeek V4 Pro (14×) |
| Document summariserCalculator: Kimi K3 · DeepSeek V4 Pro | 8,000 / 600 | 300 | $113.15no batch price | $17.53no batch price | DeepSeek V4 Pro (6.5×) |
Each row uses the same token counts for both models and list prices (for an open-weight model, the top-ranked host’s price on OpenRouter), with caching and batch discounts only where the provider publishes them. A month is 365 ÷ 12 days. The support chatbot reads 50% of its prompt from the cache. The RAG app reads 20% of its prompt from the cache. The coding agent reads 80% of its prompt from the cache. The document summariser runs as a batch job where a batch price exists. Open either model in the LLM cost calculator to change any number.
Tokenizers
Are their per-token prices comparable?
Not exactly. Each model splits text into tokens its own way, so the same prompt can be a different number of tokens on Kimi K3 and DeepSeek V4 Pro, and the cheaper per-token price isn’t always the cheaper bill. Where a tokenizer isn’t public or measured we say so rather than guess.
| Model | Tokenizer | Tokens per 1,000 words | Output per 1M words |
|---|---|---|---|
| Kimi K3 | We haven’t measured Kimi K3’s tokenizer, so its count for a given text isn’t known here. | ||
| DeepSeek V4 Pro | DeepSeek V4Our measurement | 1,145 | $0.48 |
English prose, measured on the Universal Declaration of Human Rights (2026-10-11) or taken from the provider’s published words-per-token figure. Code, JSON and other languages use more tokens per word. Count your own text in the token counter or convert with tokens to words.
FAQ
Kimi K3 vs DeepSeek V4 Pro questions
Is Kimi K3 cheaper than DeepSeek V4 Pro?
Kimi K3 costs $0.80 per million input tokens and $10 per million output tokens; DeepSeek V4 Pro costs $0.2088 and $0.4176. For a support chatbot handling 1,000 requests a day, that is about $152.46 a month on Kimi K3 against $10.24 on DeepSeek V4 Pro, so DeepSeek V4 Pro is 15× cheaper there. Try your own numbers in the LLM cost calculator.
Which has the bigger context window, Kimi K3 or DeepSeek V4 Pro?
Their context windows are about the same size. Kimi K3 reads up to 1,048,576 tokens (roughly 786,432 English words) and publishes no separate cap on reply length. DeepSeek V4 Pro reads up to 1,024,000 tokens (roughly 768,000 English words) and writes up to 384,000 tokens per reply. The window is shared between your prompt and the reply. Word counts use a rough 0.75 words per token; real ratios depend on the tokenizer.
Is Kimi K3 or DeepSeek V4 Pro better for coding?
We don’t publish benchmark scores, so this page can’t say which writes better code. What we can show is cost: a coding agent re-sending 40,000 tokens of context (80% cached) and writing 2,000 tokens, 200 times a day, costs about $267.67 a month on Kimi K3 and $18.63 on DeepSeek V4 Pro. Both support tool calling. Test both on tasks from your own repository.
Can Kimi K3 and DeepSeek V4 Pro read images?
Kimi K3 accepts images and video as well as text. DeepSeek V4 Pro accepts text only. Only Kimi K3 can read screenshots, photos or scanned pages, so for image work the choice is made for you. This is what each provider declares for its API, not a measure of how well it works.
Are Kimi K3 and DeepSeek V4 Pro open source?
Both have open weights: you can download them, run them on your own hardware, or buy them from several hosting providers. The prices on this page are typical hosted prices, so shop around. Check each model’s licence for commercial use, and see the VRAM calculator for the hardware a model needs.
Do Kimi K3 and DeepSeek V4 Pro count tokens the same way?
Not necessarily. Each model splits text into tokens with its own tokenizer, so the same prompt can be a different number of tokens on each, and per-token prices aren’t directly comparable. We haven’t measured Kimi K3’s tokenizer, so its count for a given text isn’t known here. DeepSeek V4 Pro uses about 1,145 tokens per 1,000 English words (our measurement). Count your own text with the token counter or convert with tokens to words.