DeepSeek API Pricing in 2026: The Full Cost Breakdown

DeepSeek's API is one of the cheapest ways to run frontier-class AI, and as of July 2026 the pricing is simple: DeepSeek-V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while DeepSeek-V4-Pro costs $0.435 per million input tokens and $0.87 per million output tokens (official DeepSeek pricing docs). Cache hits drop input costs by roughly 50 to 100x. That makes DeepSeek about 10 to 35 times cheaper than Western flagship models per token. The full picture is more nuanced, because sticker price is not the whole cost. Here is everything, verified against primary sources.

DeepSeek API pricing at a glance (July 2026)

All prices are per 1 million tokens, from the official pricing page:

Both models have a 1 million token context window and a 384K max output. V4-Flash allows 2,500 concurrent requests; V4-Pro allows 500.

Three things about this table are easy to miss:

• The cache-hit discount is enormous. A V4-Flash cache hit costs $0.0028 per million input tokens, which is 50x cheaper than a cache miss. If your workload reuses long prompts (system prompts, retrieved documents, codebases), your effective input cost can approach zero.
• Output costs 2x input on both models. Workloads that generate long responses (reports, code, articles) cost more than the input price suggests.
• V4-Pro's current price is a made-permanent discount. It launched at $1.74 input and $3.48 output, then DeepSeek made the promotional 75% cut permanent in late May 2026 (reported by Codersera, citing the official docs).

What changed in DeepSeek's pricing in 2026

The V4 era replaced everything you may have read about DeepSeek pricing in 2025. Three concrete changes matter:

1. The old models are going away on July 24, 2026. The legacy `deepseek-chat` and `deepseek-reasoner` model names are deprecated and will be retired on July 24, 2026, per the official change log. Until then they alias to V4-Flash's non-thinking and thinking modes. If your code still calls the old names, you have less than three weeks to migrate at the time of writing.

2. Off-peak discounts are gone. DeepSeek's 2025 off-peak program, which offered 50% off V3 and 75% off R1 during a daily 16:30 to 00:30 UTC window, officially ended on September 5, 2025 (DeepSeek V3.1 release notes). The current official pricing page lists flat rates with no time-of-day discounts. Community reports suggest peak/off-peak pricing may return for V4 in mid-July 2026 with roughly 2x pricing during Beijing business hours, but this comes from a Reddit community post, not DeepSeek's official docs, so treat it as unconfirmed until it appears on the pricing page.

3. Reasoning is now a mode, not a separate model. V4 models expose thinking and non-thinking modes under one price, instead of the separate `deepseek-reasoner` pricing of the R1 era. Budgeting is simpler, but thinking mode consumes more output tokens for the reasoning trace, so reasoning-heavy calls still cost more in practice.

DeepSeek API pricing history: how we got here

DeepSeek's pricing has moved fast, and old numbers still circulate widely. The verified timeline:

The pattern to plan around: DeepSeek changes prices and model lineups more aggressively than Western providers, usually downward, but with short migration windows. If you build on it, budget engineering time for model migrations, not just tokens.

How the caching discount actually works on your bill

DeepSeek prices cached input tokens separately, and the difference is large enough to change architecture decisions. Consider a support assistant that sends a 20,000-token system prompt and knowledge context with every request:

• Without caching, 1,000 requests a day send 20 million input tokens at $0.14 per million on V4-Flash: $2.80 per day.
• With the context cached, those same tokens cost $0.0028 per million: about $0.06 per day.

That is a 50x difference on the same workload. The practical rule: structure prompts so the stable content (instructions, documents, schemas) comes first and stays byte-identical across calls, and put the variable content (the user's question) at the end. This applies to most providers with prompt caching, but DeepSeek's cache-hit price is the most aggressive of the major APIs. If you want to model your own workload's numbers, the LLM cost calculator does this math across providers.

DeepSeek vs OpenAI vs Claude vs Gemini pricing

Per 1 million tokens, standard (non-cached) rates from each provider's official pricing page as of July 6, 2026 (OpenAI, Anthropic, Google):

What the table means in practice:

• Against flagships, DeepSeek's discount is 10 to 35x. V4-Pro's input is 11.5x cheaper than GPT-5.5 and Claude Opus 4.8; its output is 34x cheaper than GPT-5.5 and 29x cheaper than Opus.
• Against small Western models, the gap narrows to nothing. GPT-5.4-nano ($0.20/$1.25) and Gemini 2.5 Flash-Lite ($0.10/$0.40) are in V4-Flash's price band, and Flash-Lite is actually cheaper. DeepSeek's real edge is not "cheapest model available." It is frontier-class capability at budget-tier prices.
• Sonnet 5's intro pricing is temporary. Anthropic lists $2/$10 through August 31, 2026, returning to $3/$15 after. If you are budgeting past September, use the standard rate.

This is the same pattern covered in the broader enterprise AI pricing comparison: the market has split into premium flagships, aggressive mid-tier options, and near-free small models, with a 50x or larger spread between the extremes. Where a task lands on that spread is a routing decision, which is why multi-model setups beat single-model setups on cost almost by default.

Three worked cost scenarios

Token math makes the differences concrete. All numbers use the official rates above.

Scenario 1: internal support assistant. 10M input tokens and 2M output tokens per month, with 70% of input hitting DeepSeek's cache.

Scenario 2: agentic workload on a stronger model. 50M input tokens (30% cache hit) and 10M output tokens per month. Agent workloads skew input-heavy because agents re-read context across multi-step tasks.

Scenario 3: high-volume lightweight classification. 10M input, 2M output, no caching.

Scenario 3 is the honest one: for trivial tasks, Western lite models already match DeepSeek. The 20x savings appear when you need real capability, as in scenario 2. Budgeting per user instead of per token? See how much enterprise AI should cost per user.

Is DeepSeek actually good? The benchmark picture, honestly

Cheap only matters if the model performs. The evidence is genuinely mixed depending on which benchmark you trust:

• Coding benchmarks: strong. DeepSeek reports 80.6% on SWE-bench Verified for V4-Pro, the highest open-weights score and roughly tied with Gemini 3.1 Pro (SWE-bench leaderboard aggregation).
• Harder agentic benchmarks: behind the flagships. On Scale AI's contamination-resistant SWE-bench Pro, V4-Pro-Max scores 55.4% against Claude Opus 4.7 at 64.3% and GPT-5.5 at 58.6% (same source).
• Neutral composite indices: mid-tier. On the Artificial Analysis Intelligence Index, V4-Pro scores in the low 40s to low 50s depending on effort mode (mirror data), below Claude Opus 4.8 and GPT-5.5 tier scores.

A fair summary: V4-Pro is a legitimate frontier-adjacent model at a tenth of the price, excellent for coding and high-volume work, but the top Western flagships still win on the hardest reasoning and agentic tasks. Teams report the practical difference matters most for long-horizon agent runs, where small per-step quality gaps compound, a dynamic that shows up across agent orchestration platforms generally. DeepSeek's popularity is real either way: roughly 173 to 175 million cumulative app downloads and about 130 million active users by the end of 2025 (Business of Apps).

The costs that are not on the pricing page

For an individual developer, DeepSeek's API is cheap and capable, full stop. For an enterprise, four things belong in the total cost calculation:

1. Your data is stored in China. DeepSeek's own privacy policy states: "To provide you with our services, we directly collect, process and store your Personal Data in People's Republic of China" (DeepSeek privacy policy). Prompts, uploaded files, and chat history are covered. For regulated industries, that is usually disqualifying on its own.

2. Your inputs can be used for training by default. DeepSeek's terms allow using inputs and outputs (de-identified) to improve the service, with an opt-out toggle in the product (DeepSeek terms of use). Enterprises need that opt-out verified, not assumed.

3. Government restrictions keep expanding. At least 17 US states plus federal agencies including the Pentagon, NASA, the US Navy, and the Department of Commerce have banned or blocked DeepSeek on government systems, and Italy, Australia, Taiwan, and South Korea have restricted government use (The Conference Board). South Korea's privacy regulator separately found DeepSeek had transferred user prompts to a Beijing-based third party without consent (reported via WitnessAI). If your customers include government or regulated buyers, using DeepSeek's hosted API becomes a sales-cycle question, not just an infrastructure one.

4. The compliant workaround costs more than the sticker price. Because the model weights are open, you can run DeepSeek US-hosted via Amazon Bedrock or Microsoft Azure, which thousands of Bedrock customers already do (usage statistics). But hosted-cloud rates are meaningfully higher than DeepSeek's first-party API, which erodes the price advantage that brought you here. This tradeoff between cost, capability, and compliance is the central theme in choosing AI tools for enterprises with secure data.

When the DeepSeek API is the right choice

DeepSeek fits well when:

• You are a developer or startup optimizing API spend, without regulatory constraints on data location
• The workload is high-volume and cost-dominated: classification, extraction, summarization, code generation at scale
• You can structure prompts to exploit the 50x cache-hit discount
• You are building agent workflows where token volume makes flagship pricing painful, and per-step quality gaps are acceptable

It fits poorly when:

• You handle regulated, customer, or confidential data that cannot be stored in China
• You sell to government or enterprise buyers who ask vendors about their AI supply chain
• Your hardest tasks need the top of the benchmark table, where GPT-5.5 and Claude Opus still lead
• You need contractual enterprise guarantees (SOC 2 reports, DPAs, admin controls) that a raw budget API does not provide

Getting DeepSeek-level economics without the tradeoffs

The reason DeepSeek's pricing resonates is not DeepSeek itself. It is that most teams quietly overpay by routing every task, trivial or hard, to one premium flagship model. The 50x price spread between model tiers means the biggest AI cost lever is routing each task to the cheapest model that does the job well, a point the 2026 LLM pricing analysis quantifies: intelligent routing cuts spend by roughly 80% versus flagship-only usage, without the quality cliff of going budget-only.

That is the model behind Coworker: it routes each task to the right model automatically, so simple work runs on fast, cheap models and hard work runs on frontier models, at roughly 80% less than frontier API rates. Unlike a raw budget API, it is SOC 2 Type II certified and GDPR compliant with models hosted in the US, connects to 50+ enterprise tools like Salesforce, Slack, Jira, and Google Workspace, and carries organizational memory across tasks, which is the part that makes AI actually useful at work rather than another failed pilot. Pricing is public and self-serve at $29.99 per user per month with a free trial, unlike the sales-gated pricing common in enterprise AI (see what Glean costs for the contrast). If your goal is running real AI agent use cases across your tools instead of managing API keys, gateways like OpenRouter or a DIY LLM gateway solve routing but leave integration and memory to you. And if you are still mapping the landscape, start with the difference between agentic and generative AI or how teams build agents without code.

Frequently asked questions

How much does the DeepSeek API cost in 2026? DeepSeek-V4-Flash costs $0.14 per million input tokens (cache miss), $0.0028 on cache hits, and $0.28 per million output tokens. DeepSeek-V4-Pro costs $0.435 per million input tokens, $0.003625 on cache hits, and $0.87 per million output tokens. Both have 1M context windows, per the official DeepSeek pricing docs as of July 2026.

What happened to deepseek-chat and deepseek-reasoner? They are deprecated aliases that now point to DeepSeek-V4-Flash's non-thinking and thinking modes, and DeepSeek's change log says they will be retired on July 24, 2026. Code calling the old model names should migrate to the V4 names before then.

Does DeepSeek still offer off-peak discounts? No. The 2025 off-peak program (50 to 75% off during a UTC night window) ended September 5, 2025, and the current official pricing page lists flat rates. Community reports suggest peak/off-peak pricing may return for V4 in mid-July 2026, but that is unconfirmed by DeepSeek's official documentation.

How does DeepSeek pricing compare to GPT-5.5 and Claude? Per million tokens, DeepSeek-V4-Pro is roughly 11.5x cheaper than GPT-5.5 and Claude Opus 4.8 on input, and 29 to 34x cheaper on output. Against small Western models like GPT-5.4-nano or Gemini Flash-Lite, the price gap disappears, so DeepSeek's advantage is strongest where you need near-frontier capability.

Is the DeepSeek API safe for enterprise use? It depends on your data. DeepSeek's privacy policy states user data is stored in China, its terms allow training on inputs by default with an opt-out, and 17+ US states plus federal agencies have banned it on government systems. Enterprises that cannot accept those terms can run the open-weight models US-hosted on Amazon Bedrock or Azure at higher cost, or use a multi-model platform with enterprise compliance built in.

How do I get DeepSeek-level costs with enterprise compliance? Route tasks across model tiers instead of sending everything to one flagship. Multi-model routing captures most of the savings that make DeepSeek attractive, around 80% versus flagship-only usage, while keeping data with providers that offer SOC 2, US hosting, and enterprise terms. Coworker does this routing automatically across 50+ connected enterprise tools.

Related reading

• LLM cost calculator and 2026 pricing report
• Enterprise AI pricing comparison 2026
• The cheapest way to access Claude, GPT, and Gemini
• Multi-model AI vs single model
• How much should enterprise AI cost per user?
• Why enterprise AI fails