DeepSeek API Pricing vs Running It Locally: The Break-Even Point

DeepSeek is unusual: it publishes an API and releases the weights of the models behind it under a permissive licence. That raises an obvious question — if you can download it, should you run it yourself instead of paying per token? This guide gives the current API prices, the real hardware those same models need, and the arithmetic that decides it. The honest answer surprises most people.

Quick answer

• Two API models: deepseek-flash (DeepSeek-V4.1-Flash) and deepseek-v4-pro (DeepSeek-V4-Pro-0813).
• Flash is cheap: $0.15 per million input tokens and $0.60 per million output, off-peak.
• Prices halve off-peak — an unusual pricing model worth scheduling around.
• The weights are MIT-licensed and downloadable — but Flash is roughly 763B parameters and V4-Pro about 1.6 trillion. Neither is a home setup.
• Verdict: for these two models the API wins decisively. Local makes sense for the smaller distilled models, not the frontier ones.

The current prices

All figures are per one million tokens, from DeepSeek’s own pricing page, re-checked on 19 September 2026.

Both models carry a 1M-token context length and a maximum output of 384K, and both support thinking and non-thinking modes. Flash supports vision; V4-Pro does not. DeepSeek serves both over an OpenAI-format endpoint and an Anthropic-format one.

Two pricing details that matter more than the headline rate

Off-peak is half price. DeepSeek states that off-peak rates are half the peak rates, with peak hours running 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday — everything else is off-peak. That is an unusually large discount for simply scheduling work. If your workload is batch rather than interactive, shifting it outside those windows halves the bill with no code change.

Cache hits are nearly free. Input that hits the context cache costs $0.003 per million against $0.15 for a miss — a 50× difference. For anything with a long, stable system prompt or a document reused across turns, cache behaviour dominates your bill far more than the model choice does.

⚠️ One naming trap: the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but DeepSeek says those models have been retired — requests are served by V4.1-Flash and billed at the Flash price. If you are reading a tutorial that names them, it is out of date.

The part nobody checks: can you actually run these locally?

Yes, legally. Both models are published on Hugging Face under the MIT licence, which is about as permissive as it gets — our model licence checker covers what that permits. The problem is not the licence. It is the size.

Parameter counts read from each model’s safetensors index. The memory columns are plain arithmetic — parameters × bytes per weight — not benchmarks, and they cover weights only. The KV cache and runtime overhead sit on top.

To put 382 GB in context: the largest single consumer GPU on our GPU comparison holds 32 GB. You would need roughly a dozen of them, wired together, before the model loaded — and that is the small one, at 4-bit, before any context. This is data-centre hardware.

The break-even, honestly

The usual “API vs local” comparison assumes both options are available and asks which is cheaper. For DeepSeek’s frontier models that comparison never starts, because the hardware to run them costs more than most businesses spend on compute in a year — while Flash charges $0.60 per million output tokens.

Put concretely: a million output tokens is roughly 750,000 words. At off-peak Flash pricing that costs 60 cents. There is no self-hosted configuration of a 763B model that competes with that on cost alone.

So when does local win? Not on price at this scale — on the things price does not capture:

• Data that cannot leave your network. No per-token price makes an upload acceptable if the documents are confidential. That is the entire argument, and it is often decisive on its own. See self-hosted AI.
• Offline or air-gapped operation. An API needs a network; a local model does not.
• Stable costs. Hardware is a fixed cost. API pricing is set by the vendor and can change — DeepSeek’s own page reserves the right to adjust prices.
• Smaller models. This is the real opening, below.

To run the numbers against your own usage, the API pricing break-even calculator compares per-token cost against hardware over time.

What you can actually run at home

The frontier models are out of reach; DeepSeek’s smaller releases are not, and they are where local deployment makes sense.

• R1 distills — 8B, 14B, 32B and 70B variants that run on one consumer GPU. Our guide to running DeepSeek R1 locally covers the setup, and the quantization guide covers what each precision costs you.
• Larger MoE models — if you have serious memory, DeepSeek V3 vs Llama 3.3 70B covers the hardware.

Check any of them against your machine with the VRAM calculator, or browse by card size on the VRAM tier pages. For what governs generation speed once a model fits, see tokens per second.

FAQ

How much does the DeepSeek API cost?

Off-peak, deepseek-flash is $0.15 per million input tokens (cache miss) and $0.60 per million output. deepseek-v4-pro is $0.66 and $1.98. Peak rates are double.

When are DeepSeek’s peak hours?

01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. All other hours bill at the off-peak rate, which is half.

Can I run DeepSeek’s API models on my own hardware?

Legally yes — the weights are MIT-licensed. Practically no: Flash is about 763B parameters (≈382 GB even at 4-bit) and V4-Pro about 1.6 trillion. Use the distilled models instead.

Is it cheaper to self-host DeepSeek?

Not for the frontier models — the hardware cost is not close. Self-hosting wins on privacy, offline capability and cost predictability, and on smaller models where consumer hardware is enough.

What is a cache hit worth?

A great deal. Cached input is $0.003 per million against $0.15 uncached — 50× cheaper. Long reused system prompts benefit most.

Sources

• DeepSeek: Models & Pricing — prices, peak hours, context length, concurrency and model retirements, checked 19 September 2026
• Hugging Face: DeepSeek-V4.1-Flash and DeepSeek-V4-Pro — parameter counts and MIT licence

Updated 19 September 2026: all prices and both parameter counts re-verified against the sources on that date. API prices change; check the source page before budgeting. Memory figures are arithmetic from published parameter counts, not measurements.