"Free" With DeepSeek: What's Real
DeepSeek is the cheapest serious LLM lab in 2026, so "free" and "nearly free" blur together. Here's the honest breakdown of the actually-free routes and the barely-paid ones — plus one caveat you should read before sending any real data.
The Free & Nearly-Free Routes
- Consumer chat (100% free). chat.deepseek.com and the mobile app are free and unlimited for interactive use — great for testing prompts, useless for programmatic access.
- Open weights (free to self-host). DeepSeek V4 is MIT-licensed. If you have the GPUs (V4 Flash runs on ~4×24GB-class cards), you can self-host and pay nothing per token. This also sidesteps the data caveat below.
- Trial credits. New API accounts and gateways typically grant starter credit. LinkModel gives a $1 signup credit across models including
deepseek-v4-flash— enough for a lot of Flash tokens given its pricing. - Automatic caching (near-free repeats). DeepSeek's prefix caching charges cache hits at ~$0.0028/1M — a 98% discount — with no code changes. Structure prompts so the stable part comes first and repeated context is effectively free.
Why "Nearly Free" Is the Real Story
Even at full price, DeepSeek is almost free compared to peers. Verified rates (per 1M tokens):
| Model | Input | Output | Cached input |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | ~$0.0028 |
| DeepSeek V4 Pro | ~$0.44 (promo) | ~$0.87 (promo) | ~$0.004 |
V4 Flash input is ~18x cheaper than GPT-5.4 and ~36x cheaper than Claude Opus 4.8; on output — where agent loops spend most — it's dozens of times cheaper. Both models are 1M-context MoE. Full comparison in DeepSeek V4 Flash vs GPT-4o.
The Caveat: Data Residency
DeepSeek's official API processes data on servers in China. For EU/PII or regulated workloads that's a real compliance consideration. Two ways around it: self-host the open weights, or access DeepSeek through a gateway with its own data handling (LinkModel defaults to zero data retention). Also note: the legacy deepseek-chat / deepseek-reasoner aliases retire July 24, 2026 — use deepseek-v4-flash / deepseek-v4-pro.
Your First Call
DeepSeek's API is OpenAI-compatible, so it's a base-URL + model-name change from an OpenAI setup. Through a gateway you get one key across chat models:
curl -X POST https://api.linkmodel.ai/api/v1/... \
-H "Authorization: Bearer $LINKMODEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{ "model": "deepseek-v4-flash", "messages": [{"role":"user","content":"Summarize this in 3 bullets: ..."}] }'Confirm the exact chat endpoint and request schema in the docs and model catalog before wiring it up.
Bottom Line
- Just chatting/testing? The free consumer app or a $1 gateway credit.
- Want truly free at scale? Self-host the open weights.
- Building cheaply? V4 Flash at $0.14/$0.28 plus 98% cache discounts is about as close to free as production LLMs get — mind the data-residency point.
For where DeepSeek sits against everyone else, see the AI API pricing comparison and cheapest AI API guide.
Start free with a $1 credit and run V4 Flash today.
This is factual pricing/how-to information; verify current rates and data-handling terms on each provider's official pages before production use.