Short version: DeepSeek’s chat app costs nothing, the API is paid but cheap, and the model weights are open, so you can run them yourself without paying DeepSeek a cent. The longer answer depends on which surface you mean, and that’s what this guide breaks down. If you want a full walkthrough of every access path, our guide on how to use DeepSeek V4 covers web chat, API, and self-hosting in depth.
Here we’ll stick to the money question: what’s free, what isn’t, and what each option costs in August 2026, with real numbers from DeepSeek’s pricing page. And if you end up on the API path, a tool like Apidog helps you test and mock DeepSeek’s endpoints so you only pay for the calls that matter.
Is DeepSeek free? The short answer
Yes and no, depending on the surface:
- Chat (web and mobile app): free. No subscription, no Plus plan, no paywall. You sign up and chat.
- API: paid, but among the cheapest frontier-class APIs. Output tokens on the budget model cost $0.66 per million off-peak. There’s no free credit for new accounts, so you top up before your first call.
- Model weights: free to download and run. The V4 family ships under an MIT license on Hugging Face. Your only cost is hardware and electricity.
So “is DeepSeek free” has three honest answers: free for chatting, cheap for building, and free-as-in-open-source for self-hosting. The rest of this article puts numbers and limits on each.
The free chat app and its limits
The DeepSeek assistant at chat.deepseek.com and the iOS/Android apps cost nothing. That includes the reasoning mode, file uploads, and web search. Unlike ChatGPT or Claude, there’s no paid consumer plan to upsell you to; free is the only plan.
The limits are operational, not commercial. DeepSeek doesn’t publish a per-day message quota, but during traffic spikes you can hit “Server busy” responses and slower generations. There’s also no API access, no team workspace, and no uptime guarantee. For personal research, drafting, and coding questions, though, the free app covers a lot.
One more caveat: the consumer app is a hosted service in China, and your prompts are processed on DeepSeek’s servers under DeepSeek’s privacy policy. If that’s a blocker for your work, skip ahead to the local option.
DeepSeek API pricing in plain numbers
The API is not free, and as of August 16, 2026, DeepSeek bills on a peak/off-peak split. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays; every other hour is off-peak at half the peak rate. Here’s the current table, per 1M tokens:
| Model | Cache hit input | Cache miss input | Output |
|---|---|---|---|
| deepseek-v4-flash | $0.007 (off-peak) / $0.014 (peak) | $0.22 / $0.44 | $0.66 / $1.32 |
| deepseek-v4-pro | $0.022 / $0.044 | $0.66 / $1.32 | $1.98 / $3.96 |
| deepseek-v4-flash-vision-exp | $0.007 / $0.014 | $0.22 / $0.44 | $0.66 / $1.32 |
Three details matter more than the headline rates. First, context caching is automatic: repeated prompt prefixes (a fixed system prompt, few-shot examples) bill at the cache-hit rate, which is roughly 31x cheaper than a cache miss on Flash. Second, timing your batch jobs to off-peak hours halves the bill with zero code changes. Third, even at peak, V4 Flash output costs $1.32 per million tokens; comparable calls to GPT-class flagship models cost several times that.
For a concrete sense of scale: a chatbot that processes 10M input tokens (mostly cached) and 2M output tokens a month on Flash, running off-peak, lands around $1.50. That’s not free, but it’s close enough that hobby projects rarely feel it. If you’re deciding between the two models, our V4 Pro-0813 API guide covers when the flagship earns its 3x premium.
Free and near-free API routes
DeepSeek’s own platform gives new accounts no trial credit, so a truly free API key has to come from somewhere else. Your realistic options:
- OpenRouter’s free model pool. OpenRouter serves 20+ models at $0 with no credit card. The catch: DeepSeek’s free variants were pulled from the pool by mid-2026, so today OpenRouter’s DeepSeek listings are paid, starting around $0.035 per million input tokens for V4 Flash. That’s still cheaper than many first-party rates, and the 50 free requests per day on other models let you prototype the plumbing before spending anything.
- Cloud platform credits. General-purpose signup credits from cloud providers and inference hosts (the kind you get for creating an account) can be spent on hosted DeepSeek models. The models aren’t free; your credits make them free to you for a while.
- A minimal top-up on DeepSeek’s platform. Because rates are so low, $2 of balance goes a long way. It’s near-free, not free, but it’s the simplest route to the real first-party API.
We keep a regularly updated rundown of these routes in how to get a DeepSeek free API key, including which aggregators currently list no-cost variants.
Running DeepSeek locally for nothing
This is the path where “free” is literal. DeepSeek publishes its model weights on Hugging Face under an MIT license, which permits commercial use, fine-tuning, and redistribution. The V4 generation ships in two sizes: V4 Flash (284B total parameters, 13B active) and V4 Pro, whose full weights left preview in August 2026 at around 1.7T parameters.
The trade-off is hardware. V4 Flash needs roughly 33 GB of VRAM heavily quantized, or a single 80 GB H100 at FP8; V4 Pro’s full weights approach 900 GB and are a datacenter proposition. On a regular desktop, the practical route is a smaller distilled or earlier-generation model. Our walkthrough on running DeepSeek R1 locally with Ollama gets a reasoning model answering on consumer hardware in about ten minutes, no API key involved.
Local means no per-token charges, no rate limits, and no data leaving your machine. It also means you own the ops: quantization choices, serving stack (vLLM or SGLang for anything serious), and GPU costs that can exceed API spend unless your volume is high.
Chat vs API vs local: the decision in one pass
- Pick the chat app if you’re asking questions, drafting, or evaluating the model’s quality. Cost: $0. Effort: none.
- Pick the API if you’re building a product or automation. Cost: cents to a few dollars per month at hobby scale. Effort: low; the API is OpenAI-compatible, so most SDKs work by changing the base URL.
- Pick local weights if you need data privacy, offline use, or you already own GPUs. Cost: $0 in fees, real money in hardware. Effort: highest.
Most developers end up on the API, which raises a practical question: how do you build against it without burning tokens on every test run?
Test one call, mock the rest with Apidog
DeepSeek’s API follows the OpenAI format, so Apidog imports it in seconds: add the OpenAI-compatible spec, point the base URL at https://api.deepseek.com, drop your key into an environment variable, and send a request. You get the parsed response, timing, and streamed tokens in one view, which beats squinting at curl output.
The cost saver is Apidog’s mocking. Once you’ve captured one real DeepSeek response, spin up a mock server that replays realistic completions for every frontend and integration test after that. Your CI, your UI states, your error handling: all exercised against the mock at $0, while the actual API only sees the handful of calls that validate real model behavior. During peak-hour development sessions, when DeepSeek’s rates double, that discipline is the difference between a bill you notice and one you don’t.
FAQ
Is DeepSeek chat unlimited? DeepSeek publishes no message cap for the chat app, so in practice you’re limited by server capacity, not a quota. Expect occasional “Server busy” responses at high-traffic times. If you keep hitting them, several workarounds are covered in how to use DeepSeek V4 for free.
Does the DeepSeek API have a free trial? No. New platform accounts start with a zero balance and no promotional credit, so you top up before your first request. The near-free alternatives are aggregator platforms and cloud signup credits.
Is DeepSeek open source? The weights are open under an MIT license, which is what most people mean. Training data and full pipeline aren’t published, so purists call it “open weight” instead of open source. Either way, self-hosting is free and commercially permitted.
Can I use DeepSeek for coding without paying? Yes, two ways: the free chat app handles ad-hoc coding questions, and DeepSeek Harness, the company’s agentic CLI, can run against a local model at no API cost. See what DeepSeek Harness is for setup.
The bottom line
DeepSeek is free where it matters for evaluation (chat), nearly free where it matters for building (the API, at fractions of a cent per request off-peak), and genuinely free for self-hosting if you bring the hardware. Start with the chat app to judge quality, move to the API when you build, and keep your token bill boring by mocking everything you can. Download Apidog to import DeepSeek’s API, verify one real call, and mock the rest for free.
