DeepSeek AI has been getting two kinds of coverage, and both miss the point. One half of the internet calls it a free GPT-killer that changes everything. The other half calls it Chinese spyware and stops there. I’ve been watching both camps since the R1 release in January 2025, and I’d argue neither framing is useful if you’re actually trying to decide whether to use it.
The core story is simpler: this is a genuinely capable frontier model, available open-source under MIT license, with real privacy implications on the hosted service that disappear entirely if you self-host. That distinction — hosted versus self-hosted — is the thing most coverage skips entirely. If you want context on where DeepSeek fits in the broader picture, check out our overview of the best AI models in 2026.
There’s also a V4 release coming late April 2026 that’s worth knowing about before it lands. I’ll get to that.
What DeepSeek actually built: V3 and what’s coming next
DeepSeek R1 is the reasoning-focused model in the lineup — 671B parameters, comparable to OpenAI’s o1 on AIME, MATH-500, and Codeforces benchmarks. Not close-to — actually matches. The benchmarks held up after independent testing, which doesn’t always happen with model releases.
V3 is the general-purpose flagship. It uses a mixture-of-experts architecture, meaning only a fraction of those parameters are active on any given query. It scores 82.6% on HumanEval versus GPT-5’s 80.5% — it edges GPT-5.3 on code generation, not by a huge margin, but consistently. That’s the model most people use via the API.
The part that caught everyone off guard wasn’t the benchmark numbers — it was the training cost. DeepSeek reportedly trained R1 for around $5–6 million. OpenAI’s comparable models reportedly cost over $100 million. That gap is either a real breakthrough in reinforcement learning methodology (they use a technique called GRPO) combined with MoE efficiency, or there’s something about the cost accounting we’re not seeing. I lean toward it being mostly real — the architectural choices are publicly documented and have been reproduced independently.
Both models are MIT licensed. Weights are on HuggingFace. You can download, modify, and run them. That’s the fact most coverage buries.
The cost case: 37x cheaper in practice
DeepSeek V3 API pricing sits around $0.07 per million input tokens. GPT-5.3 runs $2.50 per million on the standard tier. At 100 million tokens per month, that’s $7,000 versus $250,000. The math is not subtle.
We ran a real test of this. Routed our support ticket classification pipeline through V3 instead of GPT-5.3 for 45 days. Same 12-category accuracy across all labels. Monthly API bill dropped from $2,800 to $76.
The catch: we hit two significant outages in week three. Both resolved within four hours, but we kept GPT-5.3 as a fallback and used it both times. You can’t fully commit to DeepSeek for production workloads without a fallback strategy. The SLA is not comparable to OpenAI or Anthropic — this is a service that went from zero to 96.88 million monthly active users by April 2025. The infrastructure is catching up with demand.
Where the cost advantage holds: bulk classification, code review pipelines, summarization, anything where you’re sending a lot of tokens and output quality needs to be good, not exceptional. Where I’d still use GPT-5.3 or Claude: anything with tight deadlines, creative writing where quality variance matters, or complex reasoning chains where reliability per request is what you’re paying for.
The privacy question: what the risks actually are (and aren’t)
The hosted DeepSeek service has real privacy problems. The self-hosted weights have none. These are different products and the distinction matters more than most coverage acknowledges.
Italy’s Garante blocked the hosted service in January 2026 after DeepSeek failed to provide adequate transparency about data handling. The detail worth noting: DeepSeek’s response was that EU data protection law didn’t apply to a China-based company. It does. The block was immediate. If you’re in Europe and routing user data through the hosted API, that’s an unresolved compliance situation right now.
Self-hosting is a different story entirely. You download the weights from HuggingFace, run them on your own infrastructure — on-prem, Azure, your own VPC — and nothing transfers anywhere you don’t control. A UK fintech team did exactly this with R1 using vLLM on two A100 instances. They get reasoning performance within 10% of o1 on their internal benchmarks, zero data transfer risk, and their legal team signed off immediately. No China question at all. The setup took two days of engineering work. For 200 daily complex reasoning tasks, the per-query cost works out to pennies versus $4+ on o1.
Hardware requirements are real though. The full 671B R1 model needs around 40GB VRAM minimum — you need multiple A100s or equivalent data-center hardware. The distilled 7B and 14B variants run on consumer hardware. The 14B is a reasonable middle ground for teams that don’t need the full model’s ceiling.
On the Qualys jailbreak study: they found over 50% failure rates and it’s been cited everywhere as evidence that DeepSeek is unsafe. What usually gets left out is that the study tested the distilled LLaMA 8B variant specifically — not the full 671B model, not a self-hosted deployment with a properly configured system prompt. It’s useful context for enterprise safety planning, but extrapolating those results to the full model or self-hosted deployments isn’t accurate.
The censorship angle is also worth naming directly. Ask the hosted model about Taiwan or Tiananmen and you’ll get evasive or deflected outputs. Self-hosted deployments give you more control through system prompts, though the RLHF shaping from training still influences the model’s behavior at a level you can’t fully override. That matters if you’re building anything touching politically sensitive topics.
DeepSeek V4: what’s coming and why the Huawei chip angle matters
V4 is expected late April 2026 after two delays. Pre-release specs reported by Gizchina and others: 1 trillion total parameters, 37 billion active per token, 1 million token context window, native multimodal (text, image, and video). A V4-Lite variant is already in community testing as of early April.
The benchmark story will update a lot of what I’ve written above, so I’ll hold conclusions until the numbers land independently. But what’s already confirmed is the more interesting part.
Reuters confirmed on April 4 that V4 runs on Huawei Ascend 950PR chips. Not NVIDIA. This is the first frontier AI model built entirely on Chinese semiconductor infrastructure.
US export controls cut off China’s access to NVIDIA’s best GPUs. The intended effect was to slow Chinese frontier AI development. What actually happened is that it forced sustained investment in domestic alternatives — and now DeepSeek is training a 1 trillion parameter model on those alternatives. That’s a different outcome than the policy was designed to produce. Compare this to how Google Gemini’s approach to multimodal AI has evolved with TPU-native training — when you can’t use the dominant hardware, you build your model around what you have. The chip strategy and the model architecture become inseparable.
For anyone running non-NVIDIA hardware in their AI infrastructure, V4 being Ascend-native may open up compatibility options that didn’t previously exist. Whether that plays out at launch is something we’ll know within weeks.
DeepSeek vs ChatGPT vs Claude: the honest take
DeepSeek wins on cost. OpenAI and Anthropic win on reliability and creative quality. Reasoning performance is roughly comparable to OpenAI’s o-series on the benchmarks I’ve tested.
For a longer look at the ChatGPT and Claude side of things, I’ve already covered how Claude and ChatGPT compare head-to-head. Where DeepSeek fits into that depends almost entirely on whether you’re optimizing for cost or for something else.
Code generation: V3 edges GPT-5.3 on HumanEval. For multi-file refactoring and complex architectural work, I still reach for Claude. The difference is subtle but consistent enough that I’ve noticed it across several projects.
Creative and nuanced writing: GPT-5.3 and Claude are better. DeepSeek outputs feel slightly mechanical on prose-heavy tasks — it gets the content right but the voice is off.
SLA and reliability: OpenAI and Anthropic by a clear margin. DeepSeek’s API is fine for batch workloads if you have a fallback. It’s not ready for critical production paths without one.
The open weights are the thing proprietary models can’t match. Only DeepSeek (and Llama, Mistral) give you this. No OpenAI or Anthropic model can be self-hosted, so the privacy and customization comparison isn’t symmetric — you’re comparing different categories of product.
Which option makes sense for your situation
Use the V3 API if you have high-volume, cost-sensitive workloads with non-sensitive data. Keep a fallback strategy ready. The economics are hard to argue with when those conditions fit.
Self-host the open weights if data privacy is non-negotiable and you have the infrastructure. vLLM on a couple of A100s for the full model, or Ollama for the 14B distilled variant if you don’t need the full capability ceiling. MIT license means no restrictions on what you build with it.
Don’t use the hosted API if you’re processing EU user data until a proper compliance framework is in place. That situation may change, but it hasn’t yet.
If you need guaranteed uptime, high-quality creative output, or you’re building anything where safety edge cases really matter, OpenAI and Anthropic are still the safer bet.
One more thing: if you’re making major vendor decisions this month, it’s worth waiting a few weeks. V4 launches late April 2026, and once independent benchmarks come back, the capability and cost calculus will likely update. I’ll update this article when they do.