DeepSeek Pricing

DeepSeek has become the reference point for cheap large-model inference. Buyers keep asking the same three questions: how much does the DeepSeek API actually cost per million tokens, is the chat app really free, and does the cost advantage survive contact with production reality once you add support, compliance, and reliability.

This page answers all three with real numbers pulled from DeepSeek's own pricing page and cross-checked against the current model lineup. Prices verified 2026-07-21; DeepSeek reserves the right to adjust rates, so check the DeepSeek API pricing page for the live figures before you commit contracts.

One important note on model naming: as of the July 2026 pricing update, DeepSeek's live API models are deepseek-v4-flash (the non-thinking mode) and deepseek-v4-pro (larger, higher quality). The older deepseek-chat and deepseek-reasoner aliases are scheduled for deprecation on 2026/07/24 and are being folded into the v4 lineup, so any pricing table you find that lists them separately is stale.

We use DeepSeek in production for clients where the workload is high-volume, mostly-English, and internal. It is not the right choice for every job. Read the compare section carefully before you migrate.

DeepSeek Pricing at a Glance

There are four ways to pay for DeepSeek, and they price very differently:

1. DeepSeek Chat (web and app) — free for individual consumer use. No paid subscription tier is currently advertised on deepseek.com.

2. DeepSeek API — per-token, priced in USD per 1 million tokens. Two tiers: deepseek-v4-flash for most workloads and deepseek-v4-pro for higher-quality output. Both apply a large cache-hit discount on repeated input tokens.

3. Self-hosted via open weights — no per-token fee. You pay for GPU hardware or GPU rental, plus operations. Break-even against the API depends on utilization.

4. Third-party hosts — providers like Together, Fireworks, and DeepInfra host DeepSeek's open-weight models with their own price cards. Use these when you want an alternative jurisdiction or SLA.

The API prices below are what DeepSeek charges directly. Cache-hit input pricing is roughly 50x cheaper than cache-miss input pricing on both v4-flash and v4-pro, so caching design has more impact on your bill than model choice does.

Prices verified 2026-07-21 against DeepSeek's official developer documentation.

API Pricing per Model

DeepSeek publishes prices per 1 million tokens, split between input and output, with a separate discounted rate when the input prefix hits the API's prompt cache. The numbers below are the direct-from-vendor rates as of 2026-07-21.

deepseek-v4-flash (formerly the deepseek-chat alias) — Input cache miss: $0.14 per 1M tokens. Input cache hit: $0.0028 per 1M tokens. Output: $0.28 per 1M tokens. This is DeepSeek's primary workhorse model and the price most third-party comparisons quote.

deepseek-v4-pro (larger, higher-quality tier) — Input cache miss: $0.435 per 1M tokens. Input cache hit: $0.003625 per 1M tokens. Output: $0.87 per 1M tokens. Use this when the workload actually needs stronger reasoning or code quality; do not default to it.

How to read the cache-hit line: DeepSeek automatically caches your prompt prefix. Any subsequent request that reuses that prefix (a common system prompt, a shared retrieval context, few-shot examples that do not change per call) pays the cache-hit rate on those tokens instead of the cache-miss rate. On a typical RAG or agent workload the effective input price sits somewhere between the two numbers, weighted by how much of your prompt is reused.

The older deepseek-chat and deepseek-reasoner model IDs are scheduled to be discontinued 2026/07/24. If your code still references those strings, update to the v4 aliases before the cutoff or requests will fail.

Off-Peak Discount Window

DeepSeek has historically offered a large off-peak discount on API calls during Beijing overnight hours — earlier price cards advertised roughly 50-75% off during a set nighttime window (UTC 16:30 to UTC 00:30, which is 00:30 to 08:30 China Standard Time).

As of the pricing update we verified on 2026-07-21, the current official pricing page does not display a standing off-peak discount table for the v4-flash and v4-pro tiers. That does not mean discounts are impossible — DeepSeek has run time-based promotions in the past — but there is no explicit off-peak schedule to plan around today.

Practical guidance: do not build your monthly cost model on the assumption of an off-peak rate. Price your workload at the standard cache-miss and cache-hit numbers above. If DeepSeek reintroduces a time-of-day discount, treat it as upside, not baseline. Check the DeepSeek pricing page before every contract renewal.

DeepSeek Chat (Consumer) Pricing

The DeepSeek Chat product at chat.deepseek.com and in the iOS and Android apps is free to use for individual consumers. There is no paid subscription tier equivalent to ChatGPT Plus or Claude Pro on DeepSeek's own consumer surface as of 2026-07-21.

Free-tier trade-offs: there are no published usage quotas, but the service is rate-limited and can be degraded during heavy load. Some features (file upload, extended reasoning, image input) may be gated, throttled, or region-restricted depending on the app version.

Important buyer note: the free chat product runs on infrastructure operated from mainland China and is subject to Chinese data-handling law. It is fine for personal experimentation. It is not appropriate for regulated data, client-confidential material, or anything a compliance team would want to audit. For that, use the API under a proper contract, a third-party host in a jurisdiction you accept, or a self-hosted deployment. See our DeepSeek data privacy and security risks guide for the full analysis.

Self-Hosted / Open-Weights Cost

DeepSeek's open-weight releases carry no per-token fee. Your cost is hardware plus operations. That trades a variable API bill for a fixed capacity cost, which only makes sense above a certain utilization threshold.

Hardware ballpark: the full-size DeepSeek V3 and R1 open-weight checkpoints are 671B-parameter Mixture-of-Experts models. Serving them at full precision requires multiple high-end datacenter GPUs (H100 or H200 class) in a single node — practical deployments typically use FP8 or INT8 quantization to fit onto a single 8-GPU H100/H200 box. Smaller distilled variants (7B to 70B parameters) run comfortably on one or two consumer or workstation GPUs.

Cloud-rental math: an 8x H100 node on the major GPU marketplaces rents for roughly $16-30 per hour depending on provider, commitment, and region. Run it 24/7 and you are looking at approximately $12,000-22,000 per month before egress, storage, and orchestration. Divide that by the tokens you actually generate to get your effective per-million-token cost.

When self-hosting beats the API: batch workloads, high sustained throughput, strict data-residency requirements, or a need to fine-tune on private data. When it loses: bursty traffic, low utilization, small teams, or any workload where you would end up under-using the hardware. For most teams under six figures of monthly API spend, the DeepSeek API is cheaper and less operational trouble than self-hosting.

See DeepSeek V3 Explained for the architecture and deployment detail behind these numbers.

How DeepSeek Pricing Compares

Every figure in this section comes from the vendor's own public pricing page linked in sources. Vendor pricing changes frequently — treat these as the numbers we verified on 2026-07-21 and re-check before any procurement decision.

DeepSeek v4-flash (verified 2026-07-21): $0.14 input cache miss / $0.28 output per 1M tokens.

OpenAI GPT-4o (as of 2026-07-21): approximately $2.50 input / $10.00 output per 1M tokens. GPT-4o mini: approximately $0.15 input / $0.60 output per 1M tokens. o1 and o3-class reasoning models are meaningfully more expensive on the output side. Check the OpenAI pricing page for the current live table.

Anthropic Claude Sonnet class (as of 2026-07-21): approximately $3.00 input / $15.00 output per 1M tokens. Haiku class is meaningfully cheaper. Check the Anthropic pricing page for the current lineup.

Google Gemini Pro (as of 2026-07-21): typically low single-digit dollars per 1M input tokens with tiered pricing above 128K context. Gemini Flash sits in the sub-dollar tier and is the closest direct competitor to deepseek-v4-flash on price. Check the Google AI pricing page for the current split.

Honest read: DeepSeek v4-flash is meaningfully cheaper than GPT-4o and Claude Sonnet on a straight per-token basis, and roughly comparable to GPT-4o mini and Gemini Flash. The gap widens further with cache hits, where DeepSeek's $0.0028 per 1M input tokens is among the lowest published rates in the market. It narrows or disappears once you compare against the frontier vendors' cheapest tiers on tasks their small models can actually handle.

Total Cost of Ownership for Business

Per-token pricing is not your bill. Your bill is per-token pricing multiplied by tokens per user session multiplied by users, plus caching design, plus support and compliance overhead.

Illustrative worked example — internal knowledge-base assistant, 100 employees, average 30 assisted sessions per person per week, average 4,000 input tokens (system prompt plus retrieved context) and 500 output tokens per session, 70% cache hit rate on the shared system prompt. That is roughly 12,000 sessions per week and about 52,000 sessions per month. On deepseek-v4-flash, monthly token cost lands in the low single-digit dollars to low tens of dollars range depending on exact cache-hit ratio.

The same workload on GPT-4o at $2.50 input and $10 output lands in the low hundreds of dollars per month. On Claude Sonnet at $3 and $15, higher again. On GPT-4o mini or Gemini Flash, it lands close to DeepSeek. The dollar delta at this scale is a rounding error on any real IT budget — do not migrate for it.

Where DeepSeek economics start to matter is high-volume workloads: document processing at scale, agentic loops that consume tens of millions of tokens per day, background summarization jobs, or customer-facing chat with heavy usage. At that scale the per-token gap compounds into real money, and the case for either the DeepSeek API or a self-hosted deployment strengthens.

Costs the sticker price does not show: engineering time to swap providers, evaluation cycles to confirm quality parity on your tasks, monitoring and observability for a less-mature vendor, procurement and legal review of a non-Western vendor, and the ongoing cost of maintaining a second inference path for fallback. Model those in before you decide.

Is DeepSeek Really Cheaper?

On a straight per-token basis: yes, clearly. deepseek-v4-flash is roughly one order of magnitude cheaper than Claude Sonnet and GPT-4o on output tokens, and the cache-hit input price is among the lowest published anywhere.

On a total-value basis: it depends on the workload. Frontier models still lead on complex multi-step reasoning, tool-use reliability, long-context recall past 128K, and multimodal tasks. If you swap a model that ships the job in one call for one that needs three retries, the cheaper per-token price evaporates.

Costs to weigh honestly against the per-token savings: (1) enterprise support — DeepSeek does not offer the same 24/7 incident response, dedicated account teams, or contractual SLAs that OpenAI, Anthropic, and Google sell to enterprise customers; (2) compliance — data-residency, GDPR, HIPAA, FedRAMP, and SOC 2 posture at DeepSeek is less mature and less documented than at the Western frontier labs, which matters in regulated industries; (3) reliability — third-party status pages have shown DeepSeek's API subject to capacity constraints during peak demand; (4) geopolitics — a growing number of Western buyers have policy restrictions on Chinese-hosted inference, which forces you into a third-party host or self-hosted path that erodes some of the price advantage.

Rule of thumb we apply for clients: DeepSeek wins on price when the workload is internal, high-volume, English, and the failure mode of an occasional wrong answer is acceptable. Frontier vendors win on price-plus-risk when the workload is customer-facing, regulated, low-volume, or reasoning-heavy. Most real businesses end up with both.

How to Choose the Right DeepSeek Plan

Personal experimentation, non-sensitive prompts, no budget line: use DeepSeek Chat on the web or app. Free, no signup friction, zero commitment.

Team pilot or production API workload under a few hundred million tokens per month, English text and code use cases, internal-facing: deepseek-v4-flash via the DeepSeek API. Start with a strict monthly spend cap in your dashboard.

Higher-quality reasoning, complex code generation, or workloads where v4-flash quality is measurably insufficient on your evals: deepseek-v4-pro. Do not pay for pro until you have proved flash is not enough on your actual tasks.

Regulated industry, EU data residency, or a customer contract that forbids Chinese-hosted inference: skip the DeepSeek API entirely. Either self-host the open weights on infrastructure you control, or use a Western-jurisdiction third-party host that carries DeepSeek's open-weight models under their own compliance program.

Sustained high throughput above roughly 500M-1B tokens per month with predictable load: model both the API bill and a self-hosted deployment. The crossover point moves with GPU market prices, but self-hosting can become cheaper at sustained enterprise scale.

Whichever path you pick, wire spend alerts in the DeepSeek console (or your GPU-cloud dashboard) before the first production call. That is the single control that has saved every client from a runaway prompt loop.