DeepSeek vs Kimi: which open model should you build with?

The short answer: DeepSeek for low token cost and MIT weights, Kimi for multimodal agents and app plans

Choose DeepSeek if you run many API calls and cost per token decides, or you want open weights under the MIT License: its Flash model starts at $0.15 per million input tokens off-peak. Choose Kimi if you want one model that reads images and video, a paid app with documents, slides and agents, or Kimi K3's 1M context in Claude Code. Quality differences are not established here.

Scope and evidence

Task: building with or using an open-weight Chinese frontier model, for coding agents, long documents and everyday chat. We compared the DeepSeek API models deepseek-flash and deepseek-v4-pro with Kimi K3 and K3.1 Pro on the Kimi API, the published weights on Hugging Face, and the two consumer apps. Prices are in US dollars, from DeepSeek's API pricing page, the Kimi API pricing and Kimi membership pages, checked on October 26, 2026.

This is a source-based comparison. Both companies publish benchmark tables in which each other's models appear; those are vendor-reported results, not AppFor.ai tests, and we did not run the same task on both.

Compare the same criteria

Each criterion gives DeepSeek's documented fact, Kimi's documented fact and the practical effect. Not established means the checked sources do not say.

Token price. DeepSeek: deepseek-flash costs $0.15 per million input tokens and $0.60 per million output off-peak, double that at peak hours; deepseek-v4-pro $0.66 and $1.98 off-peak. Kimi: kimi-k3 costs $3 input and $15 output, K3.1 Pro $2 and $10, with no time-of-day discount listed. What this means: for high-volume pipelines, DeepSeek's listed rates are a fraction of Kimi's.
Context and output. DeepSeek: 1M-token context and up to 384K output tokens for both API models. Kimi: 1,048,576-token context across the K3 series; maximum output is not established from the pages we checked. What this means: both handle very long inputs; check output limits before generating long files.
Vision. DeepSeek: vision supported on deepseek-flash, not on deepseek-v4-pro. Kimi: K3 is described as natively multimodal, understanding text, images and video in one model. What this means: for screenshots or video input with the larger model, Kimi documents the broader support.
Open weights. DeepSeek: weights and code for DeepSeek-V4-Pro-0813 under the MIT License; the DeepSeek-V4-Pro model card lists 1.6T total parameters with 49B active. Kimi: Kimi K3 weights under the custom Kimi K3 License; 2.8T total parameters with 104B active. What this means: read the Kimi K3 License before commercial self-hosting; DeepSeek's example vLLM setup uses a node with four GB300 GPUs.
Coding agents. DeepSeek: Anthropic- and OpenAI-format endpoints, integration guides for Claude Code, Codex and others, and its own DeepSeek Harness desktop agent. Kimi: Anthropic- and OpenAI-compatible endpoints, a Claude Code guide, and Kimi Code available on paid plans. What this means: both plug into Claude Code; which writes more acceptable patches is not established.
Consumer app. DeepSeek: a free assistant app on the web and mobile; paid app plans are not listed on the pages we checked. Kimi: a free Adagio plan plus Plus, Pro, Max and Ultra with more agent credits, documents, slides, Deep Research and scheduled tasks. What this means: DeepSeek costs nothing for chat; Kimi charges for heavier agent work.
Extra API costs. DeepSeek: Search and Fetch APIs at $0.75 per 1,000 successful requests, with tool calls through Claude Code or Codex free until billing starts on November 1, 2026, Beijing time. Kimi: built-in web search is documented, with separate pricing on its WebSearch pricing page that we did not review. What this means: agent runs that search the web add a per-call cost on DeepSeek from November.

Which fits your real work?

Classifying or summarising a large batch of text through the API. DeepSeek fits, because deepseek-flash has the lowest listed token prices here and off-peak hours halve them again. Kimi works, but at higher listed rates.

Running a coding agent on a large repository with screenshots of the UI. Kimi K3 fits, since it accepts images and video natively with a 1M context; DeepSeek fits only if you use deepseek-flash for the vision steps.

Turning research into slides and a report without writing code. Kimi's paid app plans fit, with documents, spreadsheets, slides and Deep Research listed from Plus. DeepSeek's app is free for chat, but slide creation is not listed, so it is not a supported choice for this task.

Cost of getting the work done

API, per million tokens: deepseek-flash $0.15 input and $0.60 output off-peak, $0.30 and $1.20 at peak; deepseek-v4-pro $0.66 and $1.98 off-peak, $1.32 and $3.96 at peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, except Chinese public holidays, which are off-peak in full. Kimi: kimi-k3 $3 and $15, K3.1 Pro $2 and $10, K3.1 Pro Highspeed $4 and $20, with cached input from $0.20.

Apps, monthly billing: DeepSeek's app is free. Kimi: Adagio free, Plus $19, Pro $39, Max $99 and Ultra $199 a month; with annual billing, $15, $31, $79 and $159 a month. Neither API has a subscription; both are pay as you go, and DeepSeek may adjust prices, according to its pricing page.

Kimi API pricing table for kimi-k3.1-pro, kimi-k3.1-pro-highspeed and kimi-k3 per million tokens — Kimi API model pricing page, captured 2026-10-26. Image: AppFor.ai screenshot of Moonshot AI's page. Image source ↗

Try both on the same small task

Suggested comparison, not a test we have run: send the same request through both APIs with the same reasoning effort, then compare the answer and the token usage each API reports.

The corrected function passes the test when you run it yourself.
The explanation names the real cause.
Record input and output tokens from each response and multiply by that model's listed price.
Run it at least three times per model before drawing a conclusion; one run is a signal, not a ranking.

Decision checklist

Match the conditions to your project. More detail is in our DeepSeek profile and our Kimi K3 profile.

If cost per token matters most, start with DeepSeek and schedule batch work off-peak.
If you need image or video input on the main model, start with Kimi K3.
If you will self-host commercially, compare the MIT License with the Kimi K3 License.
If you want a paid app with slides and agents, Kimi lists those plans; DeepSeek's app is free chat.
If your agents search the web, budget for DeepSeek's Search and Fetch fees from November 1.

Go to the source

Sources & further reading

Based on the official documentation below. Features and access can change; check the provider before making a decision.

Go a little further