DeepSeek API Cost for a 1M-Token Job (Verified May 2026)

A one-million-token job on DeepSeek's V4-Flash model costs $0.14 for input and $0.28 for output as of May 2026, the cheapest published rate among major LLM APIs. That headline hides the number that actually lands on your invoice, because no real job is all input or all output. We sent a real call to find out where the money goes, and on a deliberately input-light prompt, output still took 79% of the bill.

The "from $0.14" figure is the floor, not the forecast. What you pay depends on your input-to-output ratio and which model the API actually serves you, and on both counts the official docs leave room for a surprise.

DeepSeek V4-Flash API cost for a 1M-token job (verified 2026-05)

According to DeepSeek API Docs, V4-Flash is the default-tier model and the entry point for cost-sensitive work. Here is the rate card we budget against, with one correction the docs do not make obvious.

Sources: DeepSeek API Docs – Pricing and DeepSeek API Docs – Pricing Details USD, verified 2026-05-13.

We dropped the row that most pricing roundups still carry: deepseek-chat (legacy V3) at $0.27 / $1.10. When we actually called the deepseek-chat alias, the API returned deepseek-v4-flash and billed at $0.14 / $0.28. The legacy V3 rate is stale for that alias. Budget against the model you receive, not the name you typed.

We sent a real call and watched where the money went (first-hand evidence)

The fastest way to understand a 1M-token bill is to read a real invoice line, so we sent one. The full run is in our evidence pack; here is the call that reframes the headline.

We sent a deliberately input-light prompt, 41 input tokens, and the model replied with 75. That looks balanced enough that input should hold its own. It did not.

Output was 79% of a bill we engineered to favor input. The model also returned deepseek-v4-flash as the served model, confirming the alias behavior above. Scale this to a million tokens and the lesson holds: the mix decides the cost, and output carries it. A small-call latency of around 0.7 seconds, measured across repeated requests in the same evidence run, tells you the speed side of the trade is not the constraint here; price is.

What the rates mean for a real 1M-token job

A one-million-token job rarely splits evenly. The shape of your traffic sets the bill, so model two realistic shapes before you trust any single number.

A retrieval-augmented chat assistant might run 800K input tokens against 200K output across a batch of conversations. That bills $0.112 for input and $0.056 for output, $0.168 in total on V4-Flash. A code-generation workflow inverts the ratio to roughly 200K input and 800K output, billing $0.028 input and $0.224 output, $0.252 in total. Same million tokens, a 50% swing in cost, driven entirely by direction.

Blended cost is the combined input-plus-output charge for a given input-to-output ratio, and it is the figure to estimate against rather than the headline input price. Our live call is the proof in miniature: even at a near 1:1 token split, the 2× output rate pulled the bill toward output. The more your model talks back, the more output dominates.

How V4-Flash compares to alternatives at 1M tokens

The competitive question is rarely "which model" once you have settled on V4-Flash. It is "whose endpoint," and that turns out to be a compliance decision more than a price one.

According to SiliconFlow Pricing, the provider matches the official V4-Flash rate to the cent. SiliconFlow is a Chinese cloud inference provider at price parity with the official endpoint, which means routing through it buys you a different operator, not a different price. According to DeployBase, the US-routed providers tell a different story: Together AI and Azure AI Foundry run between 1.2× and 1.8× the official rate for the V3-class endpoint. That premium is the cost of US data residency, not of the model, which is the distinction most cost comparisons blur.

Choose this if you are estimating a new workload

Treat $0.14 input plus $0.28 output as the floor for any back-of-envelope estimate, then refine it with your own traffic. Sample a real day, count input and output tokens separately, multiply each by its own rate, and add them. Our two worked shapes bracket the realistic spread: the input-heavy RAG case lands near $0.168 per million job-tokens and the output-heavy code case near $0.252. If a vendor quotes you a single blended "per token" number without naming the model and splitting the two directions, treat it as incomplete, because as our live call showed, that single number cannot survive contact with a real workload.

FAQ

What does a 1M-token job actually cost on DeepSeek? On V4-Flash it ranges from $0.14 if every token is input to $0.28 if every token is output. A realistic 3:1 input-to-output mix lands near $0.175. All figures use the May 2026 official rate of $0.14 input and $0.28 output per million tokens.

Does the deepseek-chat alias still bill at the old V3 rate? No. When we called deepseek-chat on 2026-05-13 the API returned deepseek-v4-flash and billed at $0.14 / $0.28, not the legacy $0.27 / $1.10. Pricing pages that still attach the V3 rate to that alias are stale.

Is the official api.deepseek.com the cheapest source for V4-Flash? For V4-Flash, yes. SiliconFlow matches it exactly at $0.14 / $0.28, and no major third-party provider undercuts the official rate as of May 2026. US-routed providers cost more, and that premium buys data residency rather than a cheaper model.

Does DeepSeek charge minimums or commitments? No. The rate card is pay-as-you-go per token across all model tiers, with no monthly minimum or annual commitment as of May 2026.

---

This is part of the DeepSeek API pricing hub. For why output dominates the bill, see the input vs output pricing article.

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

<!-- METADATA { "title": "DeepSeek API Cost for a 1M-Token Job (May 2026)", "slug": "deepseek-1m-token-cost", "meta_description": "A 1M-token DeepSeek V4-Flash job costs $0.14 input + $0.28 output, May 2026. We sent a real call: output took 79% of the bill. Blended cost lands $0.17-$0.25.", "focus_keyword": "deepseek 1m token cost", "secondary_keywords": ["deepseek api cost per million tokens", "deepseek v4-flash rate card", "deepseek headline pricing", "deepseek per million token rate"], "tags": ["DeepSeek", "API Pricing", "Token Cost"], "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.md", "author_name": "Kevin Fan", "author_title": "Customer Success Manager", "author_linkedin": "", "author_expertise": ["Chinese LLM ecosystem", "AI infrastructure pricing"], "faq_pairs": [ {"q": "What does a 1M-token job actually cost on DeepSeek?", "a": "On V4-Flash it ranges from $0.14 if every token is input to $0.28 if every token is output. A realistic 3:1 input-to-output mix lands near $0.175. All figures use the May 2026 official rate of $0.14 input and $0.28 output per million tokens."}, {"q": "Does the deepseek-chat alias still bill at the old V3 rate?", "a": "No. When we called deepseek-chat on 2026-05-13 the API returned deepseek-v4-flash and billed at $0.14 / $0.28, not the legacy $0.27 / $1.10. Pricing pages that still attach the V3 rate to that alias are stale."}, {"q": "Is the official api.deepseek.com the cheapest source for V4-Flash?", "a": "For V4-Flash, yes. SiliconFlow matches it exactly at $0.14 / $0.28, and no major third-party provider undercuts the official rate as of May 2026. US-routed providers cost more, and that premium buys data residency rather than a cheaper model."}, {"q": "Does DeepSeek charge minimums or commitments?", "a": "No. The rate card is pay-as-you-go per token across all model tiers, with no monthly minimum or annual commitment as of May 2026."} ], "external_links_used": [ {"url": "https://api-docs.deepseek.com/quick_start/pricing/", "source_name": "DeepSeek API Docs – Pricing", "claim": "V4-Flash, V4-Pro, R1 official pricing"}, {"url": "https://api-docs.deepseek.com/quick_start/pricing-details-usd", "source_name": "DeepSeek API Docs – Pricing Details USD", "claim": "Per-model USD rate card"}, {"url": "https://www.siliconflow.com/pricing", "source_name": "SiliconFlow Pricing", "claim": "SiliconFlow matches official V4-Flash at $0.14/$0.28; lists V3.2 at $0.27/$0.42"}, {"url": "https://deploybase.ai/articles/deepseek-v3-pricing", "source_name": "DeployBase", "claim": "Together AI and Azure 1.2-1.8x markup vs official for V3-class endpoint"} ], "internal_links_used": [ {"url": "/blog/deepseek-api-pricing/", "anchor_text": "DeepSeek API pricing hub", "type": "hub"}, {"url": "/blog/deepseek-input-vs-output-pricing/", "anchor_text": "input vs output pricing article", "type": "sibling-micro"} ], "first_hand_evidence": { "source": "deepseek-api-pricing-evidence.md Finding 3 + Finding 1 + Finding 4", "measured": "41 input + 75 output tokens = $0.0000267, output 79%; deepseek-chat alias returns deepseek-v4-flash; small-call latency ~0.7s", "captured": "2026-05-13" }, "images_status": "spec-only (not generated; FAL_API_KEY unset)", "images": [ {"position": "featured", "type": "generated", "prompt": "Clean editorial chart showing a 1M-token job cost split between input and output for DeepSeek V4-Flash, with an inset of a real measured call where output took 79% of the bill. Blue input section, amber output section. Minimal background. 16:9.", "alt": "Cost breakdown chart showing a 1-million-token DeepSeek V4-Flash job split between input and output, with an inset of a live test where output accounted for 79% of the bill as of May 2026"} ] } -->