
DeepSeek chat is completely free with no subscription - the only paid product is a per-token API that's roughly 10x cheaper than frontier rivals.
DeepSeek is unusual: the chat app is completely free, and there is no consumer subscription at all. The only paid surface is the API, billed per token. In 2026 that runs from $0.14 per million input tokens on V4-Flash off-peak up to roughly $3.96 per million output tokens on V4-Pro at peak hours - still a fraction of frontier Western models. New API accounts get 5 million free tokens for 30 days.

| Surface | Cost | What you get |
|---|---|---|
| Web chat (chat.deepseek.com) | Free | Full chat on the latest V4 models, web search, file uploads; fair-use throttling at peak |
| Mobile app | Free | Same models as web chat |
| API | Per token | Pay-as-you-go against a prepaid balance; no monthly fee |
DeepSeek bills input and output separately, splits cached from uncached input, and since August 2026 charges peak rates during busy UTC windows. Indicative ranges from the official pricing page:
| Model | Input (cache hit) | Input (cache miss) | Output |
|---|---|---|---|
| V4-Flash, off-peak | ~$0.003/M | ~$0.15/M | ~$0.66/M |
| V4-Flash, peak | ~$0.006/M | ~$0.30/M | ~$1.32/M |
| V4-Pro, off-peak | ~$0.003/M | ~$0.66/M | ~$1.98/M |
| V4-Pro, peak | ~$0.006/M | ~$1.32/M | ~$3.96/M |
Source: DeepSeek API pricing, read 3 October 2026. Pro rates have run promo periods below list; the table shows the current band, not a locked price.
DeepSeek's V4 models use a mixture-of-experts architecture that activates only a slice of the model per token, so serving costs are structurally lower than dense frontier models. The result: API output that runs roughly an order of magnitude below GPT-5.x or Claude flagship rates for comparable quality on many tasks. The trade-offs are real though - peak throttling on the free chat, a prepaid balance model instead of subscriptions, and an API docs experience aimed at developers.
"Free chat plus cheap API" sounds like everything is covered, but the gap is usability: there is no consumer tier that buys you more chat, and the API requires you to build or connect something. If you want DeepSeek-quality models in a normal chat UI without managing API keys and token budgets, Krater runs DeepSeek models inside the same workspace as 500+ others - that is the route for non-developers, and we say so knowing we sell it.
Peak windows double your API bill if your jobs run during busy UTC hours - scheduling batch work off-peak is real money. Cached input is dramatically cheaper than uncached, so repeated prompts over the same documents cost far less if you structure them right. And the free chat's fair-use throttle means "unlimited" is only true at off-peak times.
Yes - the web and app chat are free with no subscription, limited only by fair-use throttling at busy times.
Per million tokens: roughly $0.15-1.32 input and $0.66-3.96 output depending on model, cache hits and peak windows.
No consumer subscription exists. The paid product is the per-token API; the chat stays free.
New API accounts get 5 million granted tokens valid for 30 days.
Yes - the free chat, or inside multi-model tools. Krater includes DeepSeek models in its standard chat plans.
DeepSeek pricing is the simplest in AI: chat is free, API is cheap, and there is nothing to subscribe to. The money question only exists for developers - and even there, off-peak scheduling and cache-aware prompting cut the bill in half. For everyone else, the cheapest DeepSeek is the free app or a workspace that includes it.