DeepSeek API Pricing 2026: Every Hosting Compared per 1M Tokens

Most pricing pages for DeepSeek tell you what the official API charges and stop there. That misses the point. Once you start shopping the same model across SiliconFlow, OpenRouter, Together AI and the official endpoint, the same-model per-million-token spread can be up to 4× (V4-Pro promo vs SiliconFlow's undiscounted rate), and the wrong choice locks a US enterprise into a data-residency posture its security team would reject on day one. We pulled live prices from every primary source on 2026-05-11, then ran the math against realistic chat workloads.

The aggregator-level answer first: DeepSeek's API charges $0.14 per million input tokens and $0.28 per million output tokens for V4-Flash, making it 35–100× cheaper than GPT-5.5 or Claude Opus 4.7 as of May 2026. That's the headline. The rest of this guide is everything you actually need to budget. Model-by-model rates, third-party markups, batch and cache discounts, the US compliance call, and the blended cost for a typical chat app.

DeepSeek API pricing at a glance (verified 2026-05)

DeepSeek operates four pricing tiers on its official endpoint: V4-Flash (the default), V4-Pro (currently 75% off through 2026-05-31), R1 (reasoning), and the legacy deepseek-chat / deepseek-reasoner pair from the V3 era. Cached input tokens are billed at one-tenth of cache-miss prices across all models, a structural discount that compounds with long system prompts.

Source: DeepSeek API Docs – Pricing and DeepSeek API Docs – Pricing Details USD, verified 2026-05-11.

The price you actually pay depends on which model you're calling and whether your traffic hits the cache. Treat the V4-Flash row as the floor for any back-of-envelope estimate and adjust upward for Pro or R1 workloads.

DeepSeek V4-Flash pricing: the default model

V4-Flash is what most production traffic should be hitting. At $0.14 per million input tokens and $0.28 per million output tokens, a 10-million-token monthly workload with a 3:1 input-to-output ratio costs roughly $1.75, and that's before cache discounts. Cache hits drop input cost to $0.0028 per million, a 50× reduction.

The 50× cache differential is not a typo. According to DeepSeek API Docs, on 2026-04-26 the platform cut cached-input prices to one-tenth of launch levels across every model. For any agent or RAG application that re-sends a fixed system prompt on every turn, this changes the cost model fundamentally. The system prompt becomes effectively free after the first call.

A practical pattern: if your prompt prefix is 8,000 tokens and your turn-specific delta is 200 tokens, the cached architecture means you pay full rate on 200 tokens per call instead of 8,200. We have seen production chat apps with long system prompts hit 70%+ cache hit rates without any code changes. DeepSeek's cache is automatic on identical prefixes.

V4-Flash is the right default unless you have specific evidence that your task needs Pro-tier reasoning. Push the heavy work to Pro selectively; serve the bulk of traffic on Flash.

DeepSeek V4-Pro pricing and the promo window

V4-Pro is currently in a promotional window: $0.435 per million input and $0.87 per million output, representing a 75% discount that expires 2026-05-31. After that date, list pricing returns: $1.74 input and $3.48 output per million. That's a 4× cost step the day the promo ends.

V4-Pro promo pricing is a vacuum-style discount: it ends without notice on 2026-05-31, after which the same API key gets billed at four times the rate for identical requests. Production teams budgeting beyond May 2026 should assume the list price. According to DeepSeek API Docs, the promo applies automatically, with no flag and no toggle. You will not get an invoice warning.

The Pro-vs-Flash decision should be evidence-based, not aspirational. Run your eval suite against both. For most chat and tool-use workloads, V4-Flash hits acceptable quality; Pro earns its premium only on tasks that genuinely benefit from the larger reasoning budget, such as complex code synthesis, multi-step planning, or dense document analysis. If your evals do not show a measurable lift, Pro is paying 3× for noise.

DeepSeek R1 pricing and when reasoning costs are justified

DeepSeek R1 is the reasoning-track model. It costs $0.55 per million input tokens (cache miss), $0.14 per million cached input, and $2.19 per million output. R1's output-to-input multiplier is roughly 4×, reflecting the model's tendency to emit long chains of thought before producing the visible answer.

R1's cost structure punishes verbose reasoning chains. A single R1 call that produces a 4,000-token reasoning trace plus a 500-token answer costs the same as eight V4-Flash calls. That math only works if R1's answer quality is materially better on your task. We recommend running both R1 and V4-Pro on your eval set and computing cost-per-correct-answer, not cost-per-call.

Use R1 when the task demands chain-of-thought: math word problems, formal verification, complex tool-use planning, multi-hop reasoning. Avoid R1 for translation, summarization, simple Q&A, and any classification task. V4-Flash is faster and 4-8× cheaper on those workloads.

DeepSeek API pricing across providers (verified 2026-05)

Four hosting paths matter for DeepSeek today: the official api.deepseek.com endpoint, SiliconFlow, OpenRouter, and Together AI. The official API is the price floor for all current-generation models. Third-party providers stack markup on top, with one specific exception worth flagging.

Sources: DeepSeek API Docs, SiliconFlow Pricing, DeployBase – DeepSeek V3 Pricing.

Two findings from this table that no single-vendor page will tell you. First: SiliconFlow charges the undiscounted V4-Pro list price ($1.74 / $3.48) while the official API is currently in a 75%-off promo. Anyone routing V4-Pro traffic through SiliconFlow today is paying 4× more than going direct. Second: OpenRouter publishes V3.2 at $0.252/$0.378, roughly 7–10% below the official api.deepseek.com V3.2 rate of $0.27/$0.42. The convenience layer does not always add cost; on this specific tier, OpenRouter undercuts the official endpoint. Their V4-Flash routing pricing is not separately documented at time of writing. Verify directly before betting on a number.

The "cheapest" provider is almost always the official API. The reasons to pay a markup are non-price: US data residency, unified billing across multiple model families, OpenAI-compatible routing layer, or rate-limit headroom on the official endpoint. For a ranked view of all Chinese LLM providers by price and performance, see our LLM API rankings.

Choose this if

• Lowest cost, no compliance constraint: official api.deepseek.com.
• US enterprise with data-residency requirements: Together AI or another US-based provider, despite the 1.5–3× markup.
• Multi-model routing layer (DeepSeek + GPT + Claude in one billing relationship): OpenRouter. Accept the markup as the cost of consolidation.

For anyone buying V4-Pro during the current promo window, the official API is the only sensible route. SiliconFlow's undiscounted Pro rate makes them the wrong choice here through 2026-05-31.

Input vs output token pricing for DeepSeek

DeepSeek charges more for output tokens than input tokens on every model, but the multiplier varies by model and matters for cost forecasting. Output token multiplier is the ratio of an API's output token price to its input token price, and it's the single number that determines whether your chat app's economics work.

For V4-Flash, output tokens cost $0.28 per million versus $0.14 for input, a 2× multiplier. For legacy deepseek-chat (V3-era), output is $1.10 versus $0.27 input, a ~4× multiplier. For R1, the multiplier is roughly 4× as well ($2.19 output vs $0.55 input). Source: DeepSeek API Docs, verified 2026-05.

The multiplier explains why migrating from legacy deepseek-chat to V4-Flash saves more than the input-price comparison suggests. Chat workloads are output-dominated when the assistant is talkative. A long-form support reply might have a 1:2 input-to-output ratio, in which case the 4×-multiplier legacy model costs disproportionately more than V4-Flash even after accounting for both rates.

Always quote the model name when comparing input/output ratios. Mixing V4-Flash and legacy V3-chat numbers in the same calculation is the most common mistake we see in third-party cost analyses.

DeepSeek batch API pricing vs streaming

DeepSeek's batch API is an inference mode where requests are queued and processed asynchronously at a discounted per-token rate, typically 50% cheaper than streaming but with minutes-to-hours latency. For workloads that don't need real-time response, such as overnight ETL, document processing, or evaluation harnesses, batch cuts the bill in half.

The trade-off is latency, not quality. Batch jobs share the same model weights as streaming; the discount comes from how DeepSeek schedules them against off-peak GPU capacity. A streaming call to V4-Flash that costs $0.14 per million input drops to roughly $0.07 per million in batch mode. For a 100-million-token evaluation pass, that's the difference between $14 and $7.

Batch is wrong for any user-facing interactive flow. If your application waits on the response before doing anything else, the latency makes batch unusable. Batch is right for bulk classification, document summarization pipelines, model evaluation runs, and any backfill task where the result needs to land within a shift, not within seconds.

Context caching: how to cut DeepSeek API costs by up to 50×

Context caching is a provider feature that stores a repeated prompt prefix server-side so repeat requests only pay for the delta tokens, reducing cost on long-system-prompt workloads. DeepSeek's cache is automatic. Identical prefixes across consecutive calls trigger the cache hit price without any flag.

The math: V4-Flash cache miss is $0.14 per million input; cache hit is $0.0028 per million. A workload with a 70% cache hit rate has a blended effective input cost of (0.7 × $0.0028) + (0.3 × $0.14) = $0.0436 per million tokens. That's a 3.2× reduction versus paying full cache-miss rates on every call, and a 50× reduction on the cached portion specifically.

For chat applications with stable system prompts (think customer support agents, coding assistants with fixed instructions), 70% cache hit rates are realistic without code changes. For one-shot APIs where every request is unique, cache is irrelevant. The differentiator is workload shape: repeated prefixes win on DeepSeek's billing model.

If your prompt engineering has the system instruction in the middle of the messages array, move it to the front. Cache hits require the prefix to match from token 0. A misplaced system prompt eats the entire cache discount silently.

Blended DeepSeek V4-Flash cost for a chat application (2026-05)

Blended price is a weighted average of input and output token costs calculated at a realistic usage ratio (e.g. 3:1 input-to-output for chat), expressed in USD per 1M tokens. For DeepSeek V4-Flash at a typical chat 3:1 input-to-output ratio without cache: (0.75 × $0.14) + (0.25 × $0.28) = $0.175 per million tokens blended.

Add the cache effect at 70% hit rate on the input portion: (0.75 × $0.0436) + (0.25 × $0.28) = $0.103 per million tokens blended. For a chat app processing 100 million tokens per month, that's $10.30 in raw API spend. The same workload on GPT-4o-class pricing runs $200–500 monthly on the input side alone, before counting output.

Numbers like these are why pricing pages fixate on the headline rate. The full economics, input vs output mix, cache hit rate, and model selection, compound into a bill that can be 10–50× cheaper than the OpenAI/Anthropic equivalent for the same conversational throughput.

The walkthrough is reproducible. Plug in your actual input-to-output ratio (measure it; don't guess), your observed cache hit rate (most teams find 50-80% on chat workloads), and the V4-Flash rate card. The blended number is what your CFO needs.

DeepSeek V4-Flash vs GPT-4o vs Claude: price per 1M tokens (as of 2026-05)

DeepSeek V4-Flash undercuts every Western frontier model by an order of magnitude on input tokens and a larger order on output tokens. The cross-reference here is third-party aggregator data; we cite NxCode's 2026 DeepSeek pricing guide for the cross-vendor table since we cannot link to direct competitors.

The "10–100× cheaper" framing is not marketing. For the same conversational chat workload, V4-Flash bills roughly 1–3% of what Claude Opus 4.7 bills and 3–5% of what GPT-4o-class models bill. Whether that price advantage is realized depends on whether V4-Flash hits acceptable quality on your task. That is the question your eval suite should answer, not your procurement spreadsheet.

US access to DeepSeek official API: status and pricing (2026-05)

The official api.deepseek.com endpoint is technically accessible from US IPs and accepts Visa/Mastercard payment from US-issued cards. There is no nationwide consumer ban. According to Ropes & Gray LLP, U.S. Navy, NASA, and Commerce Department bureaus have banned DeepSeek on government-furnished equipment. Texas and New York have extended similar bans to state-managed devices, per the same advisory.

The harder question is data residency. All personal data processed via the official API is stored on servers in China, and DeepSeek's privacy policy grants broad data exploitation rights, per ComplexDiscovery. For a startup prototype, that posture is often acceptable. For a regulated enterprise such as a defense contractor, a FedRAMP-aspiring SaaS, a HIPAA-covered health tech, or a financial services firm with cross-border data restrictions, it usually is not.

The decision tree for US enterprises: if your data is non-sensitive and your security team accepts China-stored API metadata, route through the official API for the lowest cost. If your data is sensitive or you serve regulated customers, route through a US-based provider (Together AI, Fireworks, or another CISA-friendly host) and pay the 1.5–3× markup as a compliance cost. Do not assume a VPN solves this; the data is logged at the model endpoint, not the network egress.

DeepSeek free tier: what new accounts get

New DeepSeek accounts receive 5 million free tokens, valid for 30 days, with no credit card required at signup, according to Costgoat's DeepSeek pricing guide. At V4-Flash rates, that's roughly $8.40 of credit; enough to evaluate the API thoroughly on a real workload before any spend commitment.

The 5M token grant covers both input and output across V4-Flash, R1, and legacy chat. After 30 days, the free tier disappears entirely and every token bills at standard rates. Plan your evaluation timeline accordingly: spin up the integration in week one, run the eval matrix in week two, decide on production routing in week three, and leave week four as buffer for additional benchmarks.

The free tier is not a long-term cost lever. It is an evaluation budget. Anyone treating it as a production subsidy will hit the wall at day 31 with no warning.

FAQ

Is DeepSeek API cheaper than OpenAI GPT-4o for production workloads?

DeepSeek V4-Flash at $0.14 per million input tokens is roughly 95% cheaper than GPT-4o-class input pricing (~$2.50–5.00 per million), and the output gap is larger. V4-Flash output runs at $0.28 versus GPT-4o-class at ~$10–15 per million, as verified 2026-05.

Why does DeepSeek's batch API pricing differ from its streaming API pricing?

The batch API costs roughly 50% less per token than streaming because DeepSeek schedules batch requests against off-peak GPU capacity, trading minutes-to-hours latency for the discount. As of 2026-05, batch is appropriate for ETL, evaluation, and any non-realtime workload.

Can US developers access DeepSeek's official API without restriction?

US developers can technically access api.deepseek.com from US IPs and pay with US-issued Visa or Mastercard cards. However, federal agencies (Navy, NASA, Commerce) and several states (Texas, New York) have banned DeepSeek on government devices, and all API data is stored on servers in China.

How does DeepSeek V4-Flash pricing compare to DeepSeek R1 pricing?

DeepSeek V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while R1 costs $0.55 input and $2.19 output. R1 is roughly 4× more expensive on input and 8× on output because it's the reasoning-track model designed for chain-of-thought tasks.

Which third-party provider offers the cheapest DeepSeek V3 access?

The official api.deepseek.com endpoint at $0.14 / $0.28 per million tokens (V4-Flash) is cheaper than every third-party provider as of 2026-05; SiliconFlow matches official pricing for V4-Flash but charges the undiscounted list price ($1.74 / $3.48) for V4-Pro while the official API is at 75% off.

What is the blended cost for a typical chat app using DeepSeek V4-Flash?

A chat workload with a 3:1 input-to-output ratio and 70% cache hit rate blends to roughly $0.10 per million tokens on DeepSeek V4-Flash, meaning 100 million monthly tokens cost about $10.30 in API spend.

Does the DeepSeek free tier renew?

The 5-million-token free credit applies only to new accounts, is valid for 30 days from registration, and does not renew. After the 30-day window, standard rates apply to every token across all models.

---

Methodology

Pricing data in this article was verified directly from official provider documentation (api-docs.deepseek.com, siliconflow.com/pricing) and cross-referenced against third-party aggregators (costgoat.com, pricepertoken.com, deploybase.ai) on 2026-05-11. Provider pricing changes without notice; check source URLs before billing decisions. The right provider depends on workload shape and compliance posture, not on any single price; run your eval suite and measure your actual input-to-output ratio and cache hit rate before locking a vendor.

Author: China LLM Directory Editorial. Independent Editorial Team specializing in Chinese LLM ecosystem pricing, AI infrastructure benchmarking, and cross-border AI compliance. Last verified: 2026-05-11.

<!-- METADATA { "title": "[Archive] DeepSeek API Pricing Pillar 2026 (Superseded)", "slug": "deepseek-api-pricing-pillar-archive", "meta_description": "DeepSeek V4-Flash costs $0.14/$0.28 per 1M tokens, 35-100x cheaper than GPT-5.5. Compare official API, SiliconFlow, OpenRouter pricing verified 2026-05.", "focus_keyword": "deepseek api pricing", "secondary_keywords": ["deepseek v4 pricing", "deepseek r1 pricing", "deepseek vs gpt-4o cost", "deepseek batch api pricing", "siliconflow deepseek pricing"], "tags": ["DeepSeek", "API Pricing", "LLM Cost", "Model Comparison"], "category": "Pricing", "cluster_id": "deepseek-api-pricing", "cluster_role": "pillar-archive", "author_name": "China LLM Directory Editorial", "author_title": "Independent Editorial Team", "author_linkedin": "TBD", "author_expertise": ["Chinese LLM ecosystem", "AI infrastructure pricing", "model benchmarking", "cross-border AI compliance"], "faq_pairs": [ {"q": "Is DeepSeek API cheaper than OpenAI GPT-4o for production workloads?", "a": "DeepSeek V4-Flash at $0.14 per million input tokens is roughly 95% cheaper than GPT-4o-class input pricing (~$2.50–5.00 per million), and the output gap is larger. V4-Flash output runs at $0.28 versus GPT-4o-class at ~$10–15 per million, as verified 2026-05."}, {"q": "Why does DeepSeek's batch API pricing differ from its streaming API pricing?", "a": "The batch API costs roughly 50% less per token than streaming because DeepSeek schedules batch requests against off-peak GPU capacity, trading minutes-to-hours latency for the discount. As of 2026-05, batch is appropriate for ETL, evaluation, and any non-realtime workload."}, {"q": "Can US developers access DeepSeek's official API without restriction?", "a": "US developers can technically access api.deepseek.com from US IPs and pay with US-issued Visa or Mastercard cards. However, federal agencies (Navy, NASA, Commerce) and several states (Texas, New York) have banned DeepSeek on government devices, and all API data is stored on servers in China."}, {"q": "How does DeepSeek V4-Flash pricing compare to DeepSeek R1 pricing?", "a": "DeepSeek V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while R1 costs $0.55 input and $2.19 output. R1 is roughly 4× more expensive on input and 8× on output because it's the reasoning-track model designed for chain-of-thought tasks."}, {"q": "Which third-party provider offers the cheapest DeepSeek V3 access?", "a": "The official api.deepseek.com endpoint at $0.14 / $0.28 per million tokens (V4-Flash) is cheaper than every third-party provider as of 2026-05; SiliconFlow matches official pricing for V4-Flash but charges the undiscounted list price ($1.74 / $3.48) for V4-Pro while the official API is at 75% off."}, {"q": "What is the blended cost for a typical chat app using DeepSeek V4-Flash?", "a": "A chat workload with a 3:1 input-to-output ratio and 70% cache hit rate blends to roughly $0.10 per million tokens on DeepSeek V4-Flash, meaning 100 million monthly tokens cost about $10.30 in API spend."}, {"q": "Does the DeepSeek free tier renew?", "a": "The 5-million-token free credit applies only to new accounts, is valid for 30 days from registration, and does not renew. After the 30-day window, standard rates apply to every token across all models."} ], "speakable_selectors": [".answer-capsule", "#faq"], "citable_blocks_count": 8, "internal_links_used": [ {"url": "/provider/siliconflow/", "anchor_text": "SiliconFlow", "type": "spoke", "paragraph": "intro"}, {"url": "/model/deepseek/v4-flash/", "anchor_text": "V4-Flash", "type": "spoke", "paragraph": "pricing-at-a-glance"}, {"url": "/model/deepseek/r1/", "anchor_text": "DeepSeek R1", "type": "spoke", "paragraph": "r1-pricing"}, {"url": "/provider/deepseek/", "anchor_text": "DeepSeek", "type": "spoke", "paragraph": "pricing-across-providers"}, {"url": "/rankings/", "anchor_text": "LLM API rankings", "type": "pillar", "paragraph": "pricing-across-providers"} ], "external_links_used": [ {"url": "https://api-docs.deepseek.com/quick_start/pricing/", "source_name": "DeepSeek API Docs – Pricing", "claim": "Official V4-Flash, V4-Pro, R1, and cache-hit pricing; 2026-04-26 cache price reduction"}, {"url": "https://api-docs.deepseek.com/quick_start/pricing-details-usd", "source_name": "DeepSeek API Docs – Pricing Details USD", "claim": "Legacy deepseek-chat ($0.27 input / $1.10 output) and deepseek-reasoner price details"}, {"url": "https://www.siliconflow.com/pricing", "source_name": "SiliconFlow Pricing", "claim": "SiliconFlow-hosted DeepSeek model prices including undiscounted V4-Pro at $1.74/$3.48"}, {"url": "https://deploybase.ai/articles/deepseek-v3-pricing", "source_name": "DeployBase – DeepSeek V3 Pricing", "claim": "Third-party provider markup analysis (Together AI 1.8×, Azure 20–35%, OpenRouter 2–3.5×)"}, {"url": "https://www.ropesgray.com/en/insights/alerts/2025/01/deepseek-legal-considerations-for-enterprise-users", "source_name": "Ropes & Gray LLP", "claim": "US government and state-level DeepSeek bans (Navy, NASA, Commerce, Texas, New York)"}, {"url": "https://complexdiscovery.com/when-innovation-meets-regulation-the-deepseek-privacy-controversy-and-its-compliance-fallout/", "source_name": "ComplexDiscovery", "claim": "Data-residency: all API personal data stored in China; broad data exploitation rights in privacy policy"}, {"url": "https://costgoat.com/pricing/deepseek-api", "source_name": "Costgoat", "claim": "5 million free tokens on new account registration, valid 30 days, no card required"}, {"url": "https://www.nxcode.io/resources/news/deepseek-api-pricing-complete-guide-2026", "source_name": "NxCode", "claim": "Cross-vendor pricing comparison vs GPT-4o and Claude Opus"} ], "images": [ {"position": "featured", "type": "generated", "prompt": "Professional product photography style: abstract visual of a price comparison dashboard for AI APIs, sleek dark UI with green and amber pricing tiers, soft studio lighting, 16:9 aspect ratio, clean tech aesthetic", "alt": "Price comparison dashboard showing DeepSeek V4-Flash API costs versus GPT-4o and Claude Opus across providers as of May 2026"}, {"position": "after-h2-1", "type": "stock", "query": "data center server rack", "alt": "Server infrastructure powering DeepSeek API endpoints"}, {"position": "after-h2-2", "type": "generated", "prompt": "Professional editorial illustration: clean isometric diagram showing tokens flowing into a model and being billed per million, with cache layer highlighted as a 50x cost saver, blue and teal palette, 16:9 aspect ratio, minimal background", "alt": "Token billing flow showing input, output, and cache pricing tiers for DeepSeek V4-Flash"}, {"position": "after-h2-5", "type": "stock", "query": "global network connectivity map", "alt": "Multi-provider API routing across global hosting endpoints"}, {"position": "after-h2-8", "type": "stock", "query": "cloud cost dashboard analytics", "alt": "Blended API cost dashboard for chat application workloads"}, {"position": "after-h2-10", "type": "stock", "query": "compliance audit office", "alt": "US enterprise compliance review for cross-border AI services"} ] } -->