Last Updated: 28 August 2026

Executive Summary

DeepSeek is two products wearing one name, and confusing them is the most expensive mistake a business can make here. The hosted app and API send your data to servers in China and have been banned by governments across several continents. The open weights, released under an MIT licence, can run on your own infrastructure where prompts never leave your control. The models are genuinely excellent and still among the cheapest serious AI available — though prices rose sharply on 16 August 2026. For most Western businesses the honest answer is: use the weights, not the service.

Estimated reading time: 10 minutes

Key Takeaways

• DeepSeek raised API prices on 16 August 2026, moving to peak and off-peak billing. Off-peak rates are roughly 57% above the old flat rate, and peak rates are more than triple it.
• The current models are V4-Flash and V4-Pro, both with a 1 million token context window. DeepSeek R2 has not been released, despite persistent rumours.
• The hosted service stores data on servers in China under Chinese law. Italy, South Korea, Taiwan, Australia, several US states and multiple federal agencies have restricted or banned it.
• Open weights under an MIT licence mean you can self-host, which removes the data-path risk entirely. This is the route most businesses should take.
• Most published security research tested R1 or distilled variants, not V4. The public record on the current models is thin — absence of findings is not evidence of safety.
• DeepSeek raised roughly $7.5 billion in June 2026 at a valuation near $58.9 billion, so this is no longer a scrappy outsider.

What Is DeepSeek?

DeepSeek is a Chinese AI lab, spun out of the quantitative hedge fund High-Flyer and led by founder Liang Wenfeng. It became globally famous in January 2025 when R1, a reasoning model trained at a fraction of Western costs, briefly topped the App Store and knocked hundreds of billions off US tech valuations.

Its defining choice is openness. DeepSeek publishes model weights under permissive licences — MIT for the current V4 family — which means anyone can download, inspect, modify and run them on their own hardware. That is a genuinely different proposition from OpenAI, Anthropic or Google, none of which release their frontier weights.

The 2026 lineup is the V4 family, previewed on 24 April and reaching general availability in stages: V4-Flash on 31 July and V4-Pro on 13 August. Both carry a 1 million token context window and a 384,000 token maximum output. An experimental vision model handles images.

Two persistent misconceptions are worth correcting. R2 has not been released. Despite enormous search interest and widely circulated “leaked specs,” there is no R2 in DeepSeek’s API and no official announcement, with reporting suggesting Liang held it back over performance. Anything you read about R2 benchmarks is speculation. And DeepSeek is no longer a shoestring operation — a reported $7.5 billion raise in June 2026 valued it near $58.9 billion.

Why this matters for your business: Every other tool in this series is one product. DeepSeek is two, with completely different risk profiles. The hosted API is cheap and legally fraught. The open weights are free and, self-hosted, carry none of the data-transfer risk. Deciding which one you mean is the entire evaluation.

Who Is It Best For?

DeepSeek is a strong fit for:

• Businesses with infrastructure to self-host. Running the open weights on your own servers gives frontier-class capability with no data leaving your control and no per-token bill.
• High-volume, cost-sensitive workloads. Even after August’s increase, DeepSeek remains dramatically cheaper than Western flagships for classification, extraction and bulk processing.
• Developers and technical teams. The API is OpenAI-compatible and also offers an Anthropic-compatible interface, so migration is largely a change of base URL and model name.
• Research and experimentation. Open weights under MIT mean you can fine-tune, inspect and modify without licensing negotiations.
• Businesses outside restricted sectors doing non-sensitive work. Public content generation, code scaffolding and internal drafts carry far lower stakes than customer data.

DeepSeek is probably not the right choice for:

• Any organisation handling personal, financial, health or client data through the hosted service.
• Government suppliers, defence contractors and regulated industries, where use may itself create compliance exposure.
• Businesses without engineering capacity to self-host, since the safe route requires infrastructure.
• Teams needing vendor accountability — support, SLAs and contractual data guarantees comparable to Western enterprise providers.
• Anyone needing reliable output on topics the Chinese state considers sensitive, where censorship is documented.

Key Features

V4-Flash and V4-Pro. Flash is the cost-optimised workhorse; Pro is the higher-capability flagship. Both support thinking and non-thinking modes, so one model covers routine and reasoning-heavy tasks.

1 million token context. Included at standard pricing across all V4 models, with output capped at 384,000 tokens per request. Large-document work carries no context surcharge.

Open weights under MIT. The genuinely distinctive feature. You can download V4 and run it on your own infrastructure, fine-tune it, or embed it in a product with minimal licence restrictions.

Automatic context caching. Repeated prompt prefixes are billed at a cache-hit rate roughly 30 times lower than a cache miss. For applications with a fixed system prompt, this is the single largest cost lever.

Dual API compatibility. DeepSeek exposes both an OpenAI ChatCompletions interface and an Anthropic-compatible interface, which makes swapping providers unusually straightforward in either direction.

Vision model (experimental). An image-capable variant shares Flash’s token rates, with image inputs billed as input tokens.

Mixture-of-experts efficiency. DeepSeek’s architectural work — including efficiency research published in 2026 — is a real contribution to the field, and part of why its costs are structurally lower rather than merely subsidised.

Two products, two risk profiles

                    DEEPSEEK
                        │
        ┌───────────────┴───────────────┐
        ▼                               ▼
  HOSTED SERVICE                  OPEN WEIGHTS
  app · web · API                 MIT licence
        │                               │
  Data → servers in China         Runs on your hardware
  Chinese data law applies        Prompts never leave
  Banned by multiple govts        No per-token cost
  Cheapest managed option         Needs GPU infrastructure
        │                               │
        ▼                               ▼
  Non-sensitive work only         Safe for most business use

[Illustration placeholder: DeepSeek hosted vs self-hosted risk split]

Ease of Use

For developers, DeepSeek is among the easiest APIs to adopt. It mirrors OpenAI’s interface, so an existing integration usually needs only a new base URL, API key and model name. There is a free web chat for evaluation and, for new developer accounts, a modest free token grant.

The awkwardness is administrative rather than technical. Legacy model names deepseek-chat and deepseek-reasoner were retired in 2026, so older code needs updating to the explicit deepseek-v4-flash and deepseek-v4-pro names. And since 16 August, costs vary by time of day, which means scheduling has become a cost decision rather than an operational one.

Self-hosting is a different proposition entirely. Running a 685-billion-parameter mixture-of-experts model requires serious GPU capacity, and most businesses will use a hosting provider rather than their own hardware. That is still meaningfully safer than DeepSeek’s own endpoints, because you choose the jurisdiction.

Future Signal Tip: If you use the hosted API, schedule batch work outside 01:00–04:00 and 06:00–10:00 UTC on weekdays. Off-peak rates are half of peak. For a business running overnight document processing, that scheduling decision alone halves the bill.

Performance

The models are genuinely good. DeepSeek’s mixture-of-experts approach delivers frontier-adjacent quality at a fraction of the compute, and the V4 family is competitive on coding and reasoning workloads against models costing many times more. For high-volume tasks where “good enough, very cheap” beats “best, expensive,” DeepSeek is often the rational choice.

Two performance caveats matter for business use.

The first is censorship. DeepSeek’s hosted models decline or reshape answers on topics sensitive to the Chinese state. For most commercial work this is irrelevant; for research, journalism, policy analysis or anything touching Chinese politics, it is disqualifying.

The second is that the public safety record is thinner than it looks. Widely cited studies from Cisco, Qualys and CrowdStrike tested R1 or distilled R1 variants, not V4. As of the sources reviewed here, V4 has not been independently red-teamed for jailbreak resistance or censorship behaviour. That is not evidence V4 is unsafe — it means nobody has published the test.

Integrations

DeepSeek’s integration story is compatibility rather than ecosystem. There is no marketplace, no connector library and no native CRM integration. What there is instead is a drop-in API that speaks both OpenAI’s and Anthropic’s formats, which means most existing tooling works with a configuration change.

The open weights extend this further: DeepSeek models are available through most major inference hosts and model-serving platforms, so businesses can run them via a provider in a jurisdiction they choose rather than through DeepSeek directly. For Western businesses, that is usually the correct architecture.

What you do not get is the surrounding platform — no governance layer, no audit tooling, no enterprise admin console. Everything beyond the model is your responsibility.

Pricing & Value

DeepSeek’s pricing changed materially on 16 August 2026 at 16:00 UTC, and much of the published guidance online still quotes the old rates. The current structure charges by time of day.

Peak hours run 01:00–04:00 and 06:00–10:00 UTC, on weekdays only; all other hours, including weekends, are off-peak. Off-peak V4-Flash costs $0.22 per million input tokens and $0.66 per million output, with cache hits at $0.007. Peak rates are exactly double. V4-Pro runs $0.66 and $1.98 off-peak, doubling at peak.

⚠️ This was a price rise, not a discount scheme. Before 16 August, V4-Flash was a flat $0.14 input and $0.28 output. The new off-peak rate is roughly 57% higher, and peak input is over three times the old price. Any cost model built before mid-August is now wrong.

Even so, DeepSeek remains among the cheapest serious APIs available — output on V4-Flash still undercuts most Western equivalents by a wide margin, and cache hits are close to free. A free web chat exists for evaluation, and new developer accounts have received a small token grant.

The more interesting economics are self-hosted. Open weights mean no per-token cost at all, trading a variable API bill for fixed infrastructure. For sustained high-volume workloads, that maths often favours self-hosting — and it resolves the data question at the same time.

What is your actual constraint here: cost, or where your data ends up? For most businesses the honest answer determines the deployment model more than the price does.

Strengths

• Open weights under an MIT licence, allowing self-hosting, fine-tuning and full data control.
• Among the cheapest serious AI APIs available, with cache hits near-free.
• One million token context on all V4 models with no long-context surcharge.
• Drop-in compatibility with both OpenAI and Anthropic API formats.
• Genuine architectural innovation in mixture-of-experts efficiency, not just aggressive pricing.

Limitations

• Hosted service stores data on servers in China under Chinese data and intelligence law.
• Banned or restricted by multiple governments, agencies and US states, which may create compliance exposure by itself.
• API prices rose substantially on 16 August 2026 and now vary by time of day.
• Documented censorship on topics sensitive to the Chinese state.
• No enterprise governance layer, admin console, SLA or contractual data guarantees.
• Public security testing covers older models; V4 has not been independently red-teamed.

Alternatives

Qwen or other open-weight models — the closest like-for-like if you want open weights and self-hosting. Worth comparing directly on your own workloads.

Llama or Mistral — open-weight families from US and European labs, which sidesteps the jurisdictional question entirely for buyers where that matters.

Google Gemini Flash tiers — the cheapest credible Western managed option, and the sensible comparison if you want low cost without self-hosting.

Claude Haiku or GPT’s smaller tiers — more expensive per token but with enterprise contracts, support and compliance documentation that DeepSeek does not offer.

The Future Signal

DeepSeek’s real contribution was never a product. It was a proof that frontier-class AI could be built and run far more cheaply than the industry assumed, and that the weights could be given away. That pressure is why cheap tiers from every major lab exist today.

The 2026 story is that DeepSeek is becoming a normal company. It raised billions at a large valuation, moved from flat pricing to time-of-day surge pricing, and quietly raised rates while most coverage still quoted the old numbers. The insurgent that made AI cheap is now optimising revenue like everyone else.

For business buyers, the durable lesson is architectural rather than political. DeepSeek made open weights a credible enterprise option, and open weights change the question you ask a vendor. Instead of “can I trust where your servers are?”, you can ask “can I run this where I choose?” That is a better question, and it is now available from several labs.

Watch two things: whether R2 ever ships, and whether peak pricing spreads to other providers. Surge pricing for AI inference would be a meaningful shift in how businesses budget.

Final Verdict

Overall Rating: 6.8 / 10

This score reflects DeepSeek as a business purchase, and it is pulled in two directions. The models are excellent and the open weights are genuinely valuable — that alone would score well above eight. The hosted service’s data path, the government bans, and the absence of any enterprise governance layer pull it down hard.

For technical teams that self-host, DeepSeek is one of the best value propositions in AI. For businesses that would use the hosted API with anything sensitive, it is difficult to recommend at any price.

Future Signal recommends DeepSeek’s open weights, self-hosted or run through a provider in your own jurisdiction. We do not recommend the hosted app or API for confidential, personal or client data, and organisations in regulated or government-adjacent sectors should check their obligations before using it at all.

Before you use DeepSeek — a short checklist

• Decide first: hosted service or self-hosted weights? They are different products
• Check whether your sector, clients or contracts restrict Chinese-hosted services
• Never send personal, financial, health or client data to the hosted API
• Re-price any cost model built before 16 August 2026 — rates changed
• Schedule batch work outside peak UTC windows to halve token costs
• Update legacy code from deepseek-chat and deepseek-reasoner model names
• If self-hosting, choose your inference provider’s jurisdiction deliberately
• Test for censorship on any topic material to your work before committing

Frequently Asked Questions

Is DeepSeek safe for business use? It depends entirely on which DeepSeek. The hosted app and API store data on servers in China under Chinese law, and are banned by multiple governments — avoid them for anything sensitive. The open weights, self-hosted on your own infrastructure, carry none of that data-path risk.

Has DeepSeek R2 been released? No. As of August 2026 there is no R2 in DeepSeek’s API and no official announcement, with reporting suggesting the founder held it back over performance. Any R2 benchmarks or specifications circulating are unverified speculation.

How much does DeepSeek cost now? Since 16 August 2026, pricing varies by time of day. V4-Flash is $0.22 input and $0.66 output per million tokens off-peak, doubling during peak hours. V4-Pro is $0.66 and $1.98 off-peak. Peak runs 01:00–04:00 and 06:00–10:00 UTC on weekdays.

Did DeepSeek get more expensive? Yes. V4-Flash was previously a flat $0.14 input and $0.28 output. The new off-peak rate is roughly 57% higher and peak input is more than triple the old price. Many published guides still quote the old figures.

Which governments have banned DeepSeek? Restrictions have been reported across Italy, South Korea, Taiwan, Australia and India, alongside multiple US federal agencies and states including New York, Texas and Virginia. Bans generally cover government devices and networks rather than private citizens.

Can I use DeepSeek without sending data to China? Yes, by self-hosting the open weights or running them through an inference provider in a jurisdiction you choose. This is the route most Western businesses should take, and it is the reason the open weights matter.

Does DeepSeek censor its answers? Yes, on topics sensitive to the Chinese state. For most commercial work this is irrelevant, but for research, journalism or policy analysis it is a genuine limitation worth testing before you commit.

Is the security research about DeepSeek still accurate? Partly. Widely cited studies from Cisco, Qualys and CrowdStrike examined R1 or distilled R1 variants rather than the current V4 family, and the underlying privacy findings concern the hosted consumer service. V4 has not been independently red-teamed publicly, so treat the record as incomplete rather than reassuring.

Conclusion

DeepSeek’s strength is that it broke the assumption that frontier AI has to be expensive and closed. The V4 models are genuinely capable, the pricing is still among the lowest available, and publishing the weights under an MIT licence gives businesses an option no Western frontier lab offers.

Its weakness is jurisdiction. The hosted service sends data to servers in China under laws that permit state access, and enough governments have acted on that to make its use a compliance question rather than a technical one. August’s price rise also removed some of the cost advantage that justified taking risks.

The practical conclusion is unusually clear. Treat DeepSeek as a model provider, not a service provider. Download the weights, run them where you control the data, and get most of the benefit with none of the exposure. If you must use the hosted API, restrict it to non-sensitive work, reprice your cost model against the August rates, and schedule around peak hours. That combination is genuinely worth having. The default hosted route, for most Western businesses, is not.