Contents
- Verdict
- API pricing
- What DeepSeek costs on two real automations
- A 98 percent cache discount changes how you write prompts
- Reliability on agent workloads, and what we can and cannot verify
- The self-hosting route, and why it matters here more than anywhere
- Is it actually good enough for agent work?
- How to access
- Privacy and residency
- What we build
- FAQ
- Price sources
On price per unit of capability, DeepSeek is close to unbeatable. V4-Flash lists at $0.14 per 1M input on a cache miss and $0.28 output as of August 2026, dropping to roughly $0.003 per 1M on cache hits, and third-party hosts list it lower still. It is genuinely good at agent work and the weights are MIT licensed, so you can run it yourself. The catch is not quality, it is the published data policy: DeepSeek states that data is processed and stored in the PRC and that inputs may be used for training, with an opt-out right. For non-sensitive, high-volume work that is a great trade. For customer records, use the open weights on infrastructure you control instead.
Best for
High-volume, low-sensitivity automation where token cost dominates, and self-hosted deployments where you want a capable agent model with no per-token bill at all.
Not for
Workloads carrying customer personal data, health or financial records through the first-party API, unless your compliance position genuinely accommodates PRC processing and training-by-default.
DeepSeek API pricing per 1M tokens
DeepSeek's pricing is built around the most aggressive cache discount published anywhere: a cache hit costs two percent of a cache miss, $0.0028 against $0.14 per 1M on V4-Flash. That makes it exceptionally cheap for agent loops that resend a stable system prompt. Both current models carry a 1M context window with an unusually large 384K maximum output, and both support JSON output, tool calls and Anthropic API compatibility alongside the OpenAI-shaped format. One caveat to check before you budget: DeepSeek's own pricing documentation states the API will soon adopt peak and off-peak pricing at 2x the regular rate during peak hours, defined as 09:00 to 12:00 and 14:00 to 18:00 Beijing time daily, with implementation pending official announcement. As of August 2026 it had not gone live where we could see it. Prices below are list prices as of August 2026.
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| deepseek-v4-flash (budget, agent-focused)Cache-hit input $0.0028 per 1M, a 98 percent discount | 1M in / 384K out | $0.14 | $0.28 |
| deepseek-v4-pro (flagship)Cache-hit input $0.003625 per 1M | 1M in / 384K out | $0.435 | $0.87 |
| deepseek-v4-flash via OpenRouterThird-party hosts undercut first-party list pricing; routing across several hosts | 1M | $0.084 | $0.168 |
List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.
What DeepSeek costs on two real automations
We price every provider on this site against the same two workloads. Workload A is an inbox agent: 1,500 runs a month at about 3,000 input tokens each, of which 2,400 are a stable system prompt and tool schema, plus 400 output tokens. Workload B is document extraction: 4,000 documents a month at 6,000 input and 300 output tokens each. That is 4.5M input and 0.6M output for A, and 24M input and 1.2M output for B.
On deepseek-v4-flash at $0.14 per 1M on a cache miss and $0.28 output, Workload A costs about $0.80 a month with no caching at all. Turn caching on and the 3.6M stable-prefix tokens bill at $0.0028 per 1M, which is about one cent, taking the whole workload to roughly $0.30 a month. Workload B costs about $3.70. On the flagship deepseek-v4-pro they are about $2.48 and $11.48. Through OpenRouter at $0.084 and $0.168 they are about $0.48 and $2.22, on Western infrastructure.
Compare those against the same two workloads elsewhere and the gap is the story. Workload A is about $1.62 on OpenAI's budget tier, $7.50 on Claude Haiku 4.5, $7.13 on grok-4.3 and $0.30 on DeepSeek with caching. That is not a rounding difference, it is a change in what is worth automating at all. Workflows that could never justify a flagship model, like checking every line of every invoice or summarising every call, become obviously worth doing when a run costs a hundredth of a cent.
Three caveats on the price. The announced peak-hour policy of 2x rates between 09:00 and 12:00 and 14:00 and 18:00 Beijing time is in DeepSeek's own pricing documentation as pending official announcement, so a workload scheduled in European mornings could double in cost with little warning. There is no published batch discount, so unlike OpenAI, Anthropic, Gemini and Moonshot there is no half-price overnight route. And third-party hosts on OpenRouter list V4-Flash below first-party pricing, which is unusual and worth exploiting if you want Western infrastructure anyway.
A 98 percent cache discount changes how you write prompts
Most providers give you a tenth off cached input. DeepSeek gives you a fiftieth: $0.0028 against $0.14 per 1M on V4-Flash, and $0.003625 against $0.435 on V4-Pro. At that ratio, cached tokens are close enough to free that the right design instinct inverts. Instead of trimming the system prompt to save money, you should be asking what else is worth putting in the stable prefix, because a longer, better-specified prompt with more examples costs almost nothing once it is cached and usually improves accuracy.
On Workload A the arithmetic is stark. The 3.6M monthly prefix tokens cost $0.50 uncached and about $0.01 cached. Doubling the system prompt to 4,800 tokens would add roughly a cent a month while potentially removing several classes of error. Very few cost structures reward being more thorough; this one does.
The caveat is that caching only helps when the prefix genuinely repeats. Put everything volatile at the end of the prompt: the customer name, the timestamp, the record ID, the retrieved documents. A prompt that interpolates a variable near the top gets no cache hit at all and pays the full $0.14. This is the single most common reason a DeepSeek bill comes in higher than expected.
There is no published batch API, so caching is effectively the only discount lever. That is worth weighing against providers where caching and batching stack: a cached and batched Claude Haiku job or a batched Gemini Flash-Lite job narrows the gap more than the headline rates suggest. DeepSeek still wins on price in almost every configuration, but by less than the list prices imply.
Reliability on agent workloads, and what we can and cannot verify
What is documented rather than claimed: both current models support JSON output, tool calls, chat prefix completion and an Anthropic-compatible request format alongside the OpenAI-shaped one, and V4-Flash additionally supports the Responses API. Fill-in-the-middle completion requires non-thinking mode on both. Those are the mechanics that determine whether an unattended automation parses cleanly, and they are all present.
What we will not do is quote DeepSeek's own agentic benchmark figures as fact. Vendor benchmarks select for the vendor, and we have not seen independent replication of the specific claims. Treat them as a statement of design intent, which is genuine: a budget model explicitly built for agent use rather than a shrunken chat model is a meaningful difference in what the vendor optimised for.
The published independent research applies here as everywhere. LOCA-bench, released in early 2026, grows agent context under controlled conditions and reports that most models including frontier and open-source ones degrade sharply as context grows, with the spread widening at length. A million-token window on a model this cheap is an invitation to build exactly what that benchmark says fails. Compact between steps. AgentCE-Bench, from April 2026, found that a 10 percent tool-call failure rate visibly degrades performance at every horizon, which is an argument for retry and backoff and idempotency on writes regardless of model.
The operational weak point is latency rather than intelligence. First-party latency has historically been less consistent than the US majors, and a PRC endpoint adds real network distance for a European or American business. If a customer is waiting on the response, either use a Western host running the open weights or keep DeepSeek for background steps where a few extra seconds do not matter.
The self-hosting route, and why it matters here more than anywhere
DeepSeek releases its weights under the MIT licence, which is about as permissive as software licensing gets. No revenue thresholds, no attribution requirements in your product, no restrictions on commercial use. That is a meaningfully better legal position than Meta's Llama Community Licence, which is not an OSI open-source licence and carries conditions worth a lawyer's attention at scale.
For a small business the practical question is what it takes to run. V4-Flash is a mixture-of-experts model with 284 billion total parameters and 13 billion active per token, which is the architecture that makes self-hosting plausible: you need enough memory to hold all the weights but you only pay compute for the active experts. In round numbers that is a few hundred gigabytes of GPU memory at 8-bit precision, so a four-GPU H100-class node running vLLM or SGLang, not a workstation. The flagship V4-Pro is a data-centre deployment and is not something you will run in a cupboard.
Whether self-hosting is worth it comes down to two things. If you have a genuine data-residency requirement - the automation touches records that legally cannot leave your infrastructure - then it is the only option and the cost is justified by necessity. If you are doing it purely to save money, run the numbers first: a GPU node plus the engineering time to keep it healthy will comfortably exceed a few hundred dollars a month of API spend, and API spend is someone else's uptime problem.
The middle path most small businesses should consider is a Western host running the open weights. OpenRouter, DeepInfra and similar services give you the model at low cost on infrastructure outside China, with published retention policies you can read and filter on. You lose the extreme cache-hit discount of the first-party API, but the list prices are still low and the compliance story is much easier to write down.
Is it actually good enough for agent work?
Yes, with the usual caveat that you should verify on your own cases rather than take anyone's word. The budget V4-Flash model is explicitly positioned for agentic use rather than being a scaled-down chat model, and DeepSeek's own materials report it outperforming an earlier preview of its flagship on agentic benchmarks. Vendor benchmarks are vendor benchmarks, so treat that as a signal about intent and design focus rather than a measured fact you can rely on.
What we can say from building on it is that tool calling and JSON output work as documented, the million-token context is real and useful, and the model does not have obvious pathologies in multi-step loops. Reasoning is folded into the V4 line via a thinking mode with an adjustable effort setting rather than a separate reasoning model, which simplifies model selection.
The weak points to plan around are operational rather than intellectual. First-party latency has historically been less consistent than the US majors, which matters if a customer is waiting for the response. If you are building something real-time and customer-facing, either use a Western host with better latency guarantees or keep DeepSeek for the background steps.
Our standard advice applies with extra force here: before switching a working automation, run twenty of your own real cases through both models and have someone who knows the correct answers grade the outputs. That takes an afternoon, costs almost nothing at these prices, and tells you more than any leaderboard. If you would like that evaluation run as part of a build, a free automation audit is where the conversation starts, and the automation services page sets out what a full build and handover includes.
Strengths for agent workloads
- The lowest published prices of any provider with frontier-adjacent capability, by a wide margin
- A cache-hit input discount of 98 percent, the steepest published by any provider we track, which is decisive for repetitive agent loops
- MIT-licensed open weights for both the budget and flagship models, so self-hosting is legally clean and commercially unrestricted
- A million tokens of context with a 384K maximum output, which suits document generation as well as reading
- The budget V4-Flash model is explicitly targeted at agentic use rather than being a shrunken chat model
- Both OpenAI-shaped and Anthropic-compatible request formats, so it drops into either style of agent harness
- V4-Flash is a mixture-of-experts model with 284B total and 13B active parameters, which is what makes self-hosting arithmetically plausible
Weaknesses
- Published policy states personal data is collected, processed and stored in the People's Republic of China, and that inputs are used to train and improve its models, with an opt-out right
- No published fixed retention window: the policy says data is retained as long as necessary to provide the services, which is not a number you can put in a compliance document
- No verified first-party endpoint outside China, unlike Moonshot, Z.ai and Qwen which all run international endpoints
- An announced peak-hour surcharge of 2x during Beijing working hours would materially change the economics if it goes live
- No batch API discount is published, so overnight work has no price break beyond caching
- First-party latency has historically been variable, which matters for anything customer-facing in real time
How to access DeepSeek
- Direct API at platform.deepseek.com. Cheapest first-party route, and the only one that gets you the full cache-hit discount, but it is a PRC endpoint.
- OpenRouter and other aggregators such as DeepInfra and Novita, which host the open weights on Western infrastructure. V4-Flash lists around $0.09 in and $0.18 out per 1M on OpenRouter, cheaper than first-party, and you can filter to providers with retention policies you accept.
- Self-hosting the MIT-licensed weights. V4-Flash, a mixture-of-experts model with a modest active parameter count, runs on a multi-GPU node with vLLM or SGLang. The flagship V4-Pro is a data-centre-scale deployment and is not realistic for a small business.
- The API is OpenAI-compatible in request format, so trying it against an existing integration is usually a base URL and key change.
Data privacy and residency
- DeepSeek's published privacy policy states that it directly collects, processes and stores personal data in the People's Republic of China, and that it uses inputs to improve and develop the services and to train and improve its technology, including its machine learning models. The policy also states a right to opt out of having your personal data used for training its models or optimising its technologies. This is a materially different default from OpenAI, Anthropic, Google's paid tier and Mistral, all of which state they do not train on API data by default.
- Retention is stated as a principle rather than a period: personal data is retained for as long as necessary to provide the services and for the other purposes set out in the policy. There is no fixed window equivalent to OpenAI's 30 days or Anthropic's per-feature table. If you need a retention number for a compliance document, this policy will not give you one.
- The policy does not set out comprehensive separate terms for API customers as distinct from consumer app users. Most provisions read as applying across the services. That matters, because with several other providers on this site the API defaults are materially better than the consumer ones and you can point at the document that says so.
- This is a factual difference in policy, not a claim about DeepSeek's security or intentions. Judge it the way you would judge any subprocessor: what data would actually flow, under what law would a dispute be heard, and can you explain the arrangement to a client or auditor who asks.
- For UK and EU businesses, PRC processing means an international transfer that needs a lawful basis and appropriate safeguards under UK GDPR or GDPR. That is a documentation exercise, not an impossibility, but it needs doing before the automation goes live rather than after.
- The clean way to get DeepSeek's capability without the policy question is the open weights. Running V4-Flash yourself, or on a Western host with a retention policy you have read, gives you the model without sending anything to a PRC endpoint. This is the route we recommend for anything touching customer data.
- Claims that DeepSeek operates US, Singapore or German data centres circulate widely. We could not verify any of them against an official source, so we are not repeating them as fact.
What we build on DeepSeek
- High-volume classification, tagging and routing over public or internal non-sensitive data
- Bulk content generation and rewriting where the large maximum output and low output price both help
- Self-hosted automation inside a business network, where the MIT licence means no per-token cost and no data leaving the building
- Development and testing loops, where you iterate cheaply on DeepSeek before deciding what runs in production
- The volume tier of a two-model stack, with a US flagship handling the judgement steps on sensitive data
Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.
DeepSeek FAQ
How much does the DeepSeek API cost?
As of August 2026, deepseek-v4-flash lists at $0.14 per 1M input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 output. The flagship deepseek-v4-pro is $0.435, $0.003625 and $0.87. Both carry a 1M context window and a 384K maximum output. OpenRouter lists V4-Flash at $0.084 and $0.168. DeepSeek's own documentation says peak-hour pricing at 2x is coming, pending official announcement.
What would a real DeepSeek automation cost per month?
A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $0.80 a month uncached on deepseek-v4-flash, and about $0.30 once the stable system prompt is hitting cache. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $3.70. The same two workloads on OpenAI's budget tier are roughly $1.62 and $6.24, and on Claude Haiku 4.5 roughly $7.50 and $30.00.
When will DeepSeek peak-hour pricing start?
DeepSeek's pricing documentation states the API will soon adopt peak and off-peak pricing, with peak rates at 2x the regular price between 09:00 and 12:00 and 14:00 and 18:00 Beijing time daily, and says implementation is pending official announcement. It was not live where we could observe it in August 2026. Since those hours overlap the European morning, a scheduled job could double in cost with little warning, so check before you commit to a budget.
Does DeepSeek have a batch API discount?
None is published as of August 2026. Caching is effectively the only discount lever, which matters when comparing against providers where caching and batching stack: OpenAI, Anthropic and Gemini all publish 50 percent batch discounts and Moonshot publishes 60 percent on part of its line. DeepSeek still wins on price in almost every configuration, but by less than the headline rates suggest once the other side batches.
Is DeepSeek safe to use for business data?
It depends on the data and the route. DeepSeek's published policy states data is processed and stored in the PRC and that inputs may be used for training, with an opt-out right. For public or non-sensitive internal data that may be perfectly acceptable. For customer records, the safer route is running the MIT-licensed open weights yourself or on a Western host whose retention policy you have read.
Can I self-host DeepSeek?
Yes. Both the budget and flagship V4 models are released under the MIT licence, which places no commercial restrictions on you. V4-Flash is a mixture-of-experts model that runs on a multi-GPU node with vLLM or SGLang. The flagship is a data-centre-scale deployment and is not realistic for a small business.
Is DeepSeek good enough to replace GPT for automations?
For classification, extraction, routing, bulk drafting and most agent tool-calling work, it is genuinely competitive and dramatically cheaper. Where we would still reach for a US flagship is nuanced judgement, customer-facing output that goes out unedited, and anything where the data itself rules out the first-party endpoint. A two-model stack usually beats picking one.
Does DeepSeek train on my API inputs?
Its published policy states it uses inputs to improve and develop the services and to train and improve its technology, including its machine learning models, and that you have a right to opt out of that use. That is a different default from OpenAI, Anthropic, Mistral and Google's paid tier, which all state they do not train on API data by default. The policy also does not set out comprehensively separate API terms. If training-by-default is unacceptable, use the open weights on infrastructure you control.
How long does DeepSeek keep my data?
The published policy does not give a number. It states personal data is retained for as long as necessary to provide the services and for the other purposes set out in the policy, with examples such as keeping account data while you have an account. That is a principle rather than a period, and it is weaker than OpenAI's stated 30-day abuse-monitoring window or Anthropic's per-feature retention table. If you need a retention figure for a compliance document, you will have to ask for one.
Does DeepSeek have servers outside China?
We could not verify any official first-party endpoint outside China as of August 2026. Claims about US, Singapore or German data centres circulate but we found no official source, so we are not repeating them as fact. If you need non-PRC infrastructure, use a Western host running the open weights instead.
What is the cache-hit discount and why does it matter so much?
When a request repeats a prefix DeepSeek has recently seen, that portion bills at roughly two percent of the normal input rate. Agent loops resend the same system prompt and tool schemas every turn, so on a busy agent the majority of input tokens can be cache hits. It is the steepest caching discount published by any provider we track and it is the main reason DeepSeek bills come out so low in practice.
