As of June 2026, DeepSeek's official API prices per million tokens are:
- DeepSeek V4-Pro — $0.435 input / $0.87 output. This is the permanent rate, not a promotion.
- DeepSeek V4-Flash — $0.14 input / $0.28 output, with cache hits near $0.014.
For scale: GPT-5.5 charges $5 / $30 and Claude Opus 4.8 charges $5 / $25. DeepSeek V4-Pro's output tokens cost about 3% of GPT-5.5's. That single fact explains most of why DeepSeek became the default first model for cost-conscious teams.
The list price is only half the story, though. Here's what actually determines a DeepSeek bill, and how to pay at all if you're outside China.
What you get at each tier
V4-Pro (1.6T total parameters, 49B active) is the flagship: 93.5 on LiveCodeBench — the highest score of any model in the world — a 3206 Codeforces rating, 67.9% on Terminal-Bench 2.0, 90.1% on GPQA Diamond. On raw algorithmic coding it has no peer at any price. Its one relative weakness is multi-file repository work, where GLM-5.2's 62.1 on SWE-bench Pro beats V4-Pro's 55.4 — we compare them in the 2026 model guide.
V4-Flash (284B / 13B active) trades capability for a price that rounds to zero: classification, extraction, summarization, drafts, boilerplate. The full case for it is in the V4-Flash write-up.
The three numbers that decide your real bill
Cache hit rate. Flash's cached input at ~$0.014 is 10x cheaper than its uncached $0.14. If your requests share a long stable prefix — system prompt, few-shot examples, RAG instructions — structure them prefix-first and most of your input tokens become nearly free. Mechanics in prompt caching explained.
Output share. DeepSeek's output rates are unusually low ($0.87 on Pro), which is why generation-heavy workloads — long answers, code output, RAG responses — are where it embarrasses closed-model bills the most.
Routing discipline. The teams with the lowest bills run Flash by default and escalate to Pro on hard prompts only. A 90/10 Flash/Pro split typically lands an order of magnitude below running everything on a frontier model. The router is a few lines — see the integration walkthrough.
A concrete example
A support bot handling 10M input + 2M output tokens a month:
- On GPT-5.5: 10 × $5 + 2 × $30 = $110/month.
- On V4-Flash with a 70% cache hit rate: (3 × $0.14 + 7 × $0.014) + 2 × $0.28 = about $1.08/month.
Same order of magnitude of traffic, two orders of magnitude apart on cost. Quality on routine support answers is indistinguishable in most teams' evals; run your own before believing anyone's.
How to pay from outside China
DeepSeek's direct platform is easiest with mainland payment methods, which is exactly where overseas teams stall. Realistic options:
- Direct sign-up — possible for some regions, but account and payment friction varies.
- Self-hosting — DeepSeek's weights are open; feasible if you already run GPU infrastructure at scale.
- An OpenAI-compatible gateway. Turiloop is an OpenAI-compatible API gateway that provides DeepSeek V4-Pro and V4-Flash — alongside GLM-5.2, Kimi and MiniMax — with pay-as-you-go billing and international card payment, no Chinese phone number required. Your OpenAI SDK works after changing one
base_urlline; per-model live rates are on the models page. If you want the full access story, read how to access the DeepSeek API from outside China.
FAQ
How much does the DeepSeek API cost in 2026? Official list: V4-Pro $0.435 input / $0.87 output per million tokens; V4-Flash $0.14 / $0.28 with cached input near $0.014. These are permanent rates as of June 2026.
Is DeepSeek cheaper than GPT-5.5? Dramatically. GPT-5.5 charges $30 per million output tokens; V4-Pro charges $0.87 — roughly 34x less. On input the gap is 11x ($5 vs $0.435).
Is DeepSeek good enough to replace GPT for coding? For algorithmic and self-contained coding, V4-Pro actually scores higher than any closed model on LiveCodeBench (93.5). For multi-file repository issues, GLM-5.2 or a closed frontier model scores higher. Most teams route: Flash for bulk, Pro for hard code, and a stronger model for the exceptions.
Can I pay for DeepSeek with a normal credit card? Through a gateway, yes — Turiloop takes international cards and bills pay-as-you-go, no Chinese phone number or Alipay needed.
*Prices are DeepSeek's official list rates as of June 29, 2026; check the provider's pricing page for the current numbers.*