Cheapest DeepSeek API (2026)

The cheapest way to use the DeepSeek API is the official api.deepseek.com endpoint on V4-Flash at $0.14 input and $0.28 output per million tokens, with the 50x prompt cache discount turned on; no major third party undercuts that floor, and US-routed resellers cost 1.8x to 3.5x more. That premium buys data residency, not a better model. So the real question is not which vendor is cheapest, but whether your compliance posture lets you use the cheapest one.

Who actually sells DeepSeek the cheapest (verified 2026-05)

Start with the floor. According to DeepSeek API Docs, V4-Flash on the official endpoint bills $0.14 per million input tokens and $0.28 per million output tokens, which is the lowest first-party rate DeepSeek publishes. The interesting part is what happens when you go shopping elsewhere: almost nobody beats it, and the one provider that ties it does so by passing the official rate straight through.

That tying provider is SiliconFlow. According to SiliconFlow pricing, it hosts DeepSeek V4-Flash at the same $0.14 / $0.28 as the official endpoint, so it is a genuine alternative rather than a markup. The catch sits one tier up: SiliconFlow charges the undiscounted V4-Pro list rate of $1.74 / $3.48 even while DeepSeek runs its own V4-Pro promo, roughly four times the official promotional price. If you only need V4-Flash, SiliconFlow ties the floor. If you want V4-Pro, the official endpoint is far cheaper during the promo window.

The pattern is consistent enough to state plainly: the model is the same DeepSeek weights everywhere, so a higher price is never buying you a better answer. It is buying you where the request is processed.

Why US-routed providers cost more (and when that is worth it)

According to DeployBase markup analysis, US-routed DeepSeek resellers run about 1.8x the official rate on Together AI, plus 20 to 35 percent on Azure AI Foundry, and anywhere from 2x to 3.5x on OpenRouter depending on tier. That spread is not arbitrage you can route around. It is the cost of keeping the request inside US infrastructure.

Here is the reasoning most buyers skip. The official endpoint stores all API data in mainland China, and several US government bodies (the Navy, NASA, Commerce) plus Texas and New York have banned DeepSeek on government devices. There is no nationwide consumer ban, and the official endpoint is reachable from US IPs, but a regulated US enterprise that touches the official endpoint inherits a data-residency problem. Paying 1.8x to 3.5x to a US provider is how you make that problem go away. For a hobby project or a startup with no residency obligation, that premium is pure waste. We cover the access and compliance picture in more depth in our piece on whether DeepSeek is available in the US.

So the cheapest-source decision has two branches. If you are unregulated, the floor is yours: official endpoint, V4-Flash, cache on. If you are regulated, your floor is the cheapest compliant US router, and you should treat the residency premium as a fixed line item, not a leak to plug.

First-hand evidence: how low the realized cost actually goes

Sticker prices are one thing; the bill you actually pay is another, because the cache changes the math. We called the official endpoint directly to measure it, and the cache result is the part competitor pages cannot reproduce.

We sent an identical 557-token prompt prefix twice, about a second apart, to deepseek-chat. On the second call, 512 of those 557 tokens (92 percent) came back served from cache. The input cost of that call fell from $0.0000780 cold to $0.0000077 warm, a roughly 10x drop on the input side, with no flag and no code change. According to the DeepSeek API Docs, cached input on V4-Flash bills at $0.0028 per million versus $0.14 for a miss, the 50x discount that produced that result.

One correction worth flagging, because it changes which price you should budget against. When we called the deepseek-chat alias, the API served deepseek-v4-flash, and deepseek-reasoner resolved to the same model. The legacy V3 figures some pricing pages still quote for those aliases are stale. You are getting V4-Flash at $0.14 / $0.28, so budget against that, not against a name that no longer maps to a separate model.

The cheapest realized cost, then, is a stack, not a single number: the official rate, plus V4-Flash as your default tier, plus a high prefix-cache hit rate. We break the cache mechanics down in the cache discount article.

Is the cheapest model also the cheapest per useful answer?

A lower per-token price can still lose if the model is chatty, so we ran the same prompt across three models to check. We asked each for an iterative Fibonacci function, code only. DeepSeek V4-Flash ran on its official endpoint; the two comparators were called via OpenRouter, which is the disclosure that matters here.

DeepSeek V4-Flash returned 27 input and 120 output tokens. GPT-4o, measured via OpenRouter, returned 34 in and 71 out and billed $0.000795 for that single call. Claude Sonnet 4.6, also via OpenRouter, returned 33 in and 83 out and billed $0.001344. The per-token gap between DeepSeek and the US frontier models is wide enough that even DeepSeek's longer output stays the cheaper answer by a large margin. Cheaper per token and cheaper per finished answer point the same direction here, which is not always true when a frontier model is far more concise.

FAQ

What is the absolute cheapest way to call the DeepSeek API? The official api.deepseek.com endpoint on V4-Flash at $0.14 input and $0.28 output per million tokens, with the 50x prompt cache discount active. SiliconFlow ties that floor by passing the official rate through. No major third party undercuts it.

Is SiliconFlow cheaper than the official DeepSeek API? No, it is the same. According to SiliconFlow's pricing, it hosts V4-Flash at the identical $0.14 / $0.28. It is a real alternative rather than a markup, but for V4-Pro it charges the undiscounted $1.74 / $3.48 list rate, roughly four times the official promo price.

Why do US providers like OpenRouter and Together charge more for DeepSeek? They route requests through US infrastructure, which the official endpoint does not. That premium runs about 1.8x to 3.5x depending on the provider and tier. It buys data residency, since the official endpoint stores API data in mainland China, not a better or different model.

How cheap can a single DeepSeek call really get? Very cheap when the cache fires. In our live test, an identical 557-token prefix dropped from $0.0000780 cold to $0.0000077 warm on the second call, a roughly 10x reduction on input, because 512 of 557 tokens were served from cache at the $0.0028-per-million rate.

---

This is part of the DeepSeek API pricing hub, which compares every routing option side by side.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing. Last verified: 2026-05-13.

<!-- METADATA { "title": "Cheapest Way to Use the DeepSeek API (2026)", "slug": "cheapest-deepseek-api", "meta_description": "The cheapest DeepSeek API is official V4-Flash at $0.14/$0.28 with the 50x cache discount. SiliconFlow ties it; US routers cost 1.8-3.5x. Live-tested 2026-05.", "focus_keyword": "cheapest deepseek api", "secondary_keywords": ["cheapest way to use deepseek api", "deepseek api lowest price", "cheapest deepseek provider", "deepseek api cost comparison"], "tags": ["DeepSeek", "API Pricing", "Cost Optimization"], "category": "Pricing", "cluster_id": "deepseek-api-pricing", "cluster_role": "micro", "hub_slug": "deepseek-api-pricing", "verified_until": "2026-08-11", "evidence_file": "clients/china-llm-aggregator/articles/deepseek-api-pricing-evidence.json", "author_name": "Kevin Fan", "author_title": "Customer Success Manager", "author_linkedin": "", "author_expertise": ["Chinese LLM ecosystem", "AI infrastructure pricing", "model benchmarking", "cross-border AI compliance"], "faq_pairs": [ {"q": "What is the absolute cheapest way to call the DeepSeek API?", "a": "The official api.deepseek.com endpoint on V4-Flash at $0.14 input and $0.28 output per million tokens, with the 50x prompt cache discount active. SiliconFlow ties that floor by passing the official rate through. No major third party undercuts it."}, {"q": "Is SiliconFlow cheaper than the official DeepSeek API?", "a": "No, it is the same. According to SiliconFlow's pricing, it hosts V4-Flash at the identical $0.14 / $0.28. It is a real alternative rather than a markup, but for V4-Pro it charges the undiscounted $1.74 / $3.48 list rate, roughly four times the official promo price."}, {"q": "Why do US providers like OpenRouter and Together charge more for DeepSeek?", "a": "They route requests through US infrastructure, which the official endpoint does not. That premium runs about 1.8x to 3.5x depending on the provider and tier. It buys data residency, since the official endpoint stores API data in mainland China, not a better or different model."}, {"q": "How cheap can a single DeepSeek call really get?", "a": "Very cheap when the cache fires. In our live test, an identical 557-token prefix dropped from $0.0000780 cold to $0.0000077 warm on the second call, a roughly 10x reduction on input, because 512 of 557 tokens were served from cache at the $0.0028-per-million rate."} ], "external_links_used": [ {"url": "https://api-docs.deepseek.com/quick_start/pricing/", "source_name": "DeepSeek API Docs – Pricing", "claim": "V4-Flash official rate $0.14/$0.28; cached input $0.0028/M vs $0.14 miss = 50x"}, {"url": "https://siliconflow.com/", "source_name": "SiliconFlow Pricing", "claim": "Hosts V4-Flash at $0.14/$0.28 matching official; V4-Pro at undiscounted $1.74/$3.48"}, {"url": "https://deploybase.ai/", "source_name": "DeployBase markup analysis", "claim": "Together AI ~1.8x official; Azure AI Foundry +20-35%; OpenRouter 2x-3.5x"} ], "internal_links_used": [ {"url": "/blog/deepseek-api-pricing/", "anchor_text": "DeepSeek API pricing hub", "type": "hub"}, {"url": "/blog/is-deepseek-available-in-us/", "anchor_text": "DeepSeek is available in the US", "type": "sibling-micro"}, {"url": "/blog/deepseek-cache-discount/", "anchor_text": "cache discount article", "type": "sibling-micro"} ], "first_hand_evidence": { "source": "deepseek-api-pricing-evidence.json runs: cache_cold, cache_warm, coding_comparison_same_prompt, alias_deepseek-chat, alias_deepseek-reasoner", "measured": "557-token prefix call #2 cache_hit=512 of 557 (92%); input cost $0.0000780 cold -> $0.0000077 warm (~10x); deepseek-chat and deepseek-reasoner both served deepseek-v4-flash; Fibonacci same-prompt: V4-Flash 27in/120out, GPT-4o via OpenRouter 34in/71out billed $0.000795, Claude Sonnet 4.6 via OpenRouter 33in/83out billed $0.001344", "captured": "2026-05-13" }, "images_status": "spec-only (not generated; FAL_API_KEY unset)", "images": [ {"position": "featured", "type": "generated", "prompt": "Clean editorial comparison chart ranking DeepSeek API sources by realized cost: official api.deepseek.com and SiliconFlow at the $0.14/$0.28 floor, then US-routed providers stacked 1.8x to 3.5x above, annotated with the residency-premium label. Blue and amber palette, minimal background, 16:9.", "alt": "Comparison chart ranking DeepSeek API providers by cost, with official api.deepseek.com and SiliconFlow at the floor and US-routed providers priced 1.8x to 3.5x higher for data residency"} ] } -->