In the Weeds: What DeepSeek V4 Actually Means If You're Not a Developer

Last Thursday, a lab in Hangzhou released a piece of software that can hold a conversation, write code, solve calculus problems, and reason through legal briefs — and it costs roughly one-tenth of what the same work costs from OpenAI. The lab is DeepSeek. The software is called V4. And unless you read AI newsletters for fun, you probably didn't hear about it at all.

That's fine. You don't need to use DeepSeek V4. You don't need to know how it works under the hood. But you should know it exists, because the ripple effects are going to touch the AI tools you already use — and the price you pay for them — within months.

This is the In the Weeds column, where we take one technical development and strip the jargon until you can explain it at dinner. Today: what DeepSeek V4 actually means for someone who has never opened a terminal in their life.

First, the headline numbers

DeepSeek V4 has 1.6 trillion parameters. That number is staggering on paper, but it doesn't tell you much by itself — it's like saying a car has 847 horsepower. Impressive, but what matters is how it drives.

Here's what matters: DeepSeek V4 Flash, the lightweight version built for everyday tasks, costs $0.14 per million input tokens. For comparison, GPT-4o — the model behind most ChatGPT conversations — runs about $2.50 per million input tokens on the API. That's roughly eighteen times more expensive.

If you're a person who pays $20 a month for ChatGPT Plus and that's all you've ever done, the per-token pricing doesn't touch you directly. But it touches the people who build the tools you use. Every AI-powered app — the ones that summarize your emails, write your cover letters, help your kid with algebra — pays for AI by the token behind the scenes. When the cost of that intelligence drops by a factor of ten, the math changes for every company in the stack.

Cheaper ingredients mean cheaper meals. Not overnight. But inevitably.

What "open source" means here (and what it doesn't)

DeepSeek released V4 as open source, which is a phrase that gets thrown around loosely enough to be confusing. Here's what it actually means in this context.

Open source, in traditional software, means the code is public. Anyone can read it, modify it, redistribute it. Linux is open source. Firefox is open source. WordPress is open source. The idea is that when anyone can inspect the machinery, the machinery gets better, and nobody's locked in to one vendor.

When an AI lab says a model is "open source," they usually mean the model weights are publicly available. The weights are the learned knowledge — billions of numbers that represent what the model learned during training. If you have the weights, you can run the model on your own hardware. You can fine-tune it for your specific needs. You can build products on top of it without paying the original lab a cent.

DeepSeek V4's weights are available under a permissive license. Companies, researchers, and tinkerers can download them and run the model themselves.

What "open source" does NOT mean: you don't get the training data. You don't get the exact recipe for how they trained it. You don't get the reinforcement-learning-from-human-feedback pipeline. You get the finished product, not the kitchen.

For you, the non-developer, this matters for one practical reason: competition. When a powerful model is open, dozens of companies race to build products on top of it. That race drives prices down and quality up. The same way Android being open source meant you didn't have to buy an iPhone to get a smartphone, open-source AI models mean you don't have to use one company's chatbot to get AI help.

The specialist hospital (what "mixture of experts" means)

DeepSeek V4 uses an architecture called Mixture of Experts, often abbreviated MoE. This is the key to how it's so cheap to run despite being so large.

Think of it this way. Imagine a hospital with one doctor — a brilliant generalist. Every patient who walks in sees that one doctor, who has to know cardiology, neurology, pediatrics, oncology, dermatology, everything. That doctor works extremely hard on every single case. This is how most AI models work: the entire model activates for every query.

Now imagine a different hospital. This one has 256 specialists. When you walk in with chest pain, you don't see all 256 of them. The triage nurse routes you to the cardiologist and maybe the pulmonologist. Two specialists work your case. The other 254 sit idle.

That second hospital is a Mixture of Experts model.

DeepSeek V4 has 1.6 trillion total parameters, but for any given question, it only activates about 52 billion of them. The "triage nurse" — technically called the gating network — decides which experts are relevant and routes the query to them. The result: you get the depth of a massive model with the cost of a much smaller one.

This is why V4 Flash costs $0.14 per million tokens. You're not paying to wake up the entire hospital. You're paying for the two specialists who actually worked your case.

A brief history, because context matters

DeepSeek didn't come out of nowhere, even if it feels that way.

The company was founded in 2023 as a research subsidiary of High-Flyer Capital Management, a Chinese quantitative hedge fund. Their first notable release, DeepSeek Coder, was a code-focused model that surprised people by competing with GitHub Copilot on certain benchmarks. Useful, but niche.

Then came DeepSeek V2 in mid-2024. This was the model that got the industry's attention — not because it was the best at anything, but because it was shockingly cheap to run. V2 introduced the Mixture of Experts architecture (more on that above) to a model that was genuinely useful for everyday tasks. Overnight, companies that were paying $30 per million tokens for GPT-4 could get 80% of the quality for $0.28. That pricing forced every other lab to respond.

V3 followed in late 2024 with improvements in reasoning and a longer context window. It was a solid update, but incremental. The kind of release that matters to benchmarks but not to dinner conversations.

V4 is different. V4 isn't incremental. It's the first open-source model that credibly matches the best proprietary models across a wide range of tasks, while costing a fraction of the price. That combination — quality parity at radically lower cost — is what makes it significant to anyone, not just developers.

The backstory matters because it tells you something about DeepSeek's strategy. They're not trying to build the flashiest chatbot or win a consumer subscription war. They're trying to be the cheapest high-quality infrastructure. They want to be the commodity that everyone builds on top of. And with V4, they're closer to that goal than they've ever been.

What V4 is genuinely good at

Every model has a personality, even if nobody at the lab intended it. DeepSeek V4 has strengths that are distinct and measurable.

Math and formal reasoning. On AIME 2025 (a benchmark of competition-level math problems), V4 scored 75.4%. That's competitive with the best models from any lab. If your use case involves calculations, proofs, logic puzzles, or structured problem-solving, V4 is strong. This isn't abstract — it means that if you're a parent helping your kid with precalculus and you're out of your depth, V4 can walk through the problem step by step with the kind of rigor that a math teacher would. (If that's your situation, 📚The Study Buddy pairs well with any model for structured tutoring.)

Code. DeepSeek has always been code-forward — the company started as an offshoot of a quantitative hedge fund, and their first public model was DeepSeek Coder. V4 continues that tradition. On LiveCodeBench, it outperforms most open-source alternatives and holds its own against proprietary models. For non-developers, this matters less directly — but it means tools that need to process data, automate workflows, or generate reports are getting access to a cheaper engine.

Long conversations. V4 supports a 128K-token context window — roughly 300 pages of text. You can paste in an entire contract, a book chapter, a semester's worth of lecture notes, and ask questions about it without the model losing the thread. This is a practical feature, not a benchmark curiosity. If you've ever pasted a long document into ChatGPT and gotten a response that clearly forgot what was on page one, you've hit a context-window limit. V4's window is large enough that most real-world documents fit comfortably.

Instruction following. V4 is notably better than its predecessors at doing what you actually asked. If you say "give me three bullet points and nothing else," you're more likely to get three bullet points and nothing else. This sounds trivial, but inconsistent instruction following is one of the most common frustrations with AI tools. When a tool adds unsolicited caveats, ignores your formatting request, or answers a different question than the one you asked, that's an instruction-following failure. V4 has fewer of them.

What V4 is not good at (and what to use instead)

Honest assessments are more useful than cheerleading. Here's where V4 falls short.

No image generation. DeepSeek V4 is a text model. It doesn't create images, edit photos, or generate visual content. If you need that, 🖼️The AI Image Tool Matcher can point you to the right tool for your specific use case — whether that's Midjourney for aesthetic work, ChatGPT Images for text-heavy graphics, or something else entirely. (We published an honest review of Midjourney V8.1 this week, if you're in the market.)

Conversational polish. GPT-4o and Claude have a warmth and conversational flow that DeepSeek V4 doesn't quite match. V4's responses can feel more utilitarian — accurate, but less fun to read. If you're using AI as a thinking partner or a writing collaborator, the personality gap matters.

Data privacy questions. DeepSeek is based in Hangzhou, China. If you use their hosted API (as opposed to running the open-source model yourself), your data flows through Chinese servers. For casual use — brainstorming, studying, generating ideas — this is probably fine. For sensitive business data, medical information, or legal documents, it's worth thinking about. This isn't scaremongering; it's the same calculus you'd apply to any cloud service based in any country with different data-governance laws than yours.

The good news: because V4 is open source, companies outside China can (and already do) run it on their own infrastructure. The model's strengths without the jurisdictional questions.

Multimodal input. V4 can't "see" images you share with it the way GPT-4o or Claude can. No uploading a photo of your kid's math homework and asking "what's wrong here?" For that workflow, you still need a multimodal model.

The price war nobody's talking about

Here's the part that actually affects your life.

In 2024, running a million tokens through GPT-4 cost roughly $30 on the input side. By early 2025, GPT-4o brought that down to about $2.50. Now DeepSeek V4 Flash sits at $0.14.

That's a 200x cost reduction in eighteen months.

When something gets 200 times cheaper, the entire economy around it rearranges. Think about what happened when phone calls went from dollars-per-minute to essentially free. People didn't just make the same calls for less money — they started making different calls. Businesses that couldn't afford call centers suddenly could. Products that required real-time voice communication became possible. The entire category of "things you can do with a phone call" expanded.

The same thing is starting to happen with AI inference. Tools that were too expensive to build at GPT-4 pricing become viable at DeepSeek V4 pricing. A startup that wants to give every user a personal AI tutor — not a luxury product at $50/month, but a free feature — can now afford the compute. A small business that wants AI to process every customer email, not just flag the important ones, can now justify the cost.

This doesn't mean your ChatGPT subscription drops to $2 next month. OpenAI, Anthropic, and Google have premium features, brand trust, and ecosystem lock-in that justify their pricing. But the floor has dropped. And when the floor drops, even the premium players have to respond.

We're already seeing it. OpenAI launched GPT-4o Mini. Anthropic launched Claude 3.5 Haiku. Google launched Gemini Flash. Every major lab now has a "fast and cheap" tier that didn't exist two years ago. DeepSeek didn't cause all of that, but DeepSeek — starting with V2 in 2024 — demonstrated that "fast and cheap" didn't have to mean "mediocre." That changed the conversation.

What this means for the tools you already use

If you use AI tools regularly — and if you're reading a-gnt, you probably do — here's the practical impact.

Tools built on top of open-source models will get cheaper. Many AI-powered apps don't use OpenAI or Anthropic at all. They use open-source models like Llama, Mistral, or DeepSeek, hosted on their own servers or through cloud providers. When a better open-source model drops, these tools upgrade their backend and pass the quality improvement (and sometimes the cost savings) to you.

"AI features" will show up in more places. The biggest barrier to adding AI features to a product has been cost. At $0.14 per million tokens, it becomes economically feasible to add AI to tools where it would have been prohibitively expensive before. Expect more AI-powered features in your note-taking apps, email clients, and project management tools — not because every product needs AI, but because the cost barrier to experimenting is nearly gone.

Quality will converge. This is the most important trend. Two years ago, GPT-4 was clearly better than everything else, and you paid a premium for that quality. Today, the gap between the best proprietary model and the best open-source model is small and shrinking. For most tasks a normal person does — writing emails, studying, brainstorming, summarizing documents — the difference between DeepSeek V4 and GPT-4o is marginal. The tools you use will increasingly swap models behind the scenes based on cost and speed, and you won't notice.

Privacy-conscious alternatives will improve. Because V4 is open source, privacy-focused companies can run it on local hardware. Tools that promise "your data never leaves your device" used to mean "you get a worse model." That trade-off is fading.

Should you use DeepSeek directly?

Probably not, unless you have a specific reason.

If you're happy with ChatGPT, Claude, or Gemini, there's no compelling reason to switch to DeepSeek's chatbot. The advantages of V4 are mostly upstream — they benefit the builders who use it as infrastructure, and those benefits trickle down to you through the tools they build.

But if you're curious, DeepSeek's chat interface at chat.deepseek.com is free and functional. It's worth trying for a few queries just to calibrate your sense of what "open source AI" feels like in 2026. The 🔍DeepSeek Translator on a-gnt is a good starting prompt — it's designed to help you ask DeepSeek questions in a way that plays to its strengths.

For math and science questions, V4 is genuinely excellent. If your teenager is grinding through AP Calculus and wants a second tutor, or if you're a small business owner trying to sanity-check financial projections, the reasoning capability is real. Pair it with 📚The Study Buddy for a more structured study experience, or check out our piece on building a homework routine your kid won't hate for a framework that works with any AI tool.

The geopolitics, briefly

We can't talk about DeepSeek without talking about the fact that it's Chinese. Not because that makes it sinister, but because it adds context that matters.

The U.S. has restricted the export of advanced AI chips (specifically NVIDIA's A100 and H100 GPUs) to China since October 2022. Those restrictions were supposed to slow China's AI development. DeepSeek trained V4 on hardware that predates the restrictions, combined with clever engineering that reduces the compute needed per parameter. The fact that they produced a world-class model despite the chip embargo is, depending on your perspective, either an impressive feat of engineering or a cautionary tale about export controls.

For you, the person reading this article, the geopolitical dimension matters only if you use DeepSeek's hosted services. If your data flows through DeepSeek's servers in China, Chinese data-governance laws apply. China's data-security laws allow the government to request access to data stored on Chinese servers by Chinese companies. This is not hypothetical — it's the legal framework.

The practical takeaway: for low-stakes use (brainstorming, studying, generating ideas, creative writing), using DeepSeek's chat interface is no more risky than using any foreign cloud service. For high-stakes use (sensitive business data, medical records, legal strategy, financial details), either use a non-Chinese hosting provider running the open-source weights, or use a different model entirely.

This isn't unique to DeepSeek. The same logic applies to any AI service based in any jurisdiction with different privacy laws than your own. The difference is that DeepSeek is the first Chinese model good enough that ordinary people might actually want to use it, so the question is newly relevant.

The bigger picture

Every few months, a new model drops and the AI commentariat declares it a "GPT killer" or a "paradigm shift." Most of the time, the truth is more boring and more important: the overall quality floor rises, the cost drops, and the tools you use get quietly better.

DeepSeek V4 isn't going to replace ChatGPT on your phone. But it's going to make the next wave of AI tools cheaper, faster, and more accessible. The startup that couldn't afford to build an AI tutor last year can now afford it. The nonprofit that wanted AI-powered translation for refugees can now justify the cost. The solopreneur who's been paying $100/month across three AI subscriptions might find that one of them drops its price or adds features it couldn't afford before. (If that solopreneur is you, Your First AI Week is a day-by-day plan for getting the most out of the tools that exist right now.)

That's what a price-performance breakthrough actually looks like from the outside. Not a single product you switch to, but a tide that lifts every product in the category.

The best thing you can do with this information isn't to go sign up for DeepSeek. It's to know that when a tool you already like announces it's getting faster, cheaper, or more capable — and when a new tool shows up that does something nobody could afford to build before — there's a reason. A lab in Hangzhou released their work to the world, and the world is building on it.

That's what "open source" means in practice. Not a download link. A permission slip.

Quick reference: DeepSeek V4 at a glance

What it is. A 1.6-trillion-parameter language model from Chinese AI lab DeepSeek, released April 24, 2026.

What it costs. V4 Flash: $0.14/million input tokens. V4 full: $0.80/million input tokens. Both are roughly 10-18x cheaper than equivalent GPT-4o pricing.

What it's good at. Math, code, long documents, structured reasoning, following instructions precisely.

What it's not good at. Image generation (none), visual input (none), conversational warmth (functional but dry), multimedia tasks.

Privacy note. Hosted version runs on Chinese servers. Open-source version can run anywhere.

Should you switch? No — unless you have a specific technical or cost reason. The benefits will reach you through the tools built on top of it.

Want to try it? 🔍The DeepSeek Translator on a-gnt is a good starting point — it's a prompt designed to help you get the most out of DeepSeek's strengths.

For more on the AI tools your family already uses (and some they should), check out The AI Tools Your Teenager Is Already Using and The One-Person Business: Your First AI Week.