DeepSeek V4 Explained: Models, Pricing and Open Source
DeepSeek V4 drew attention for near top level results at low prices and an open license. It is worth understanding what it actually is, what it is not, and how its pricing has changed since launch.
This guide covers the models, features, benchmarks, pricing, and limits based on DeepSeek’s own documentation and independent reviews.
What Is DeepSeek V4
DeepSeek V4 is a family of open weight language models from the Chinese lab DeepSeek, released April 24, 2026 in two versions, Pro and Flash. Both support a 1 million token context window by default across DeepSeek’s official services, along with separate Thinking and Non-Thinking modes, and both are compatible with OpenAI and Anthropic API formats, making them easy to drop into existing tooling.
Pro vs Flash
| Model | Reported Size | Notes |
|---|---|---|
| V4 Pro | About 1.6 trillion total parameters, about 49 billion active | Larger, stronger model, trained on 33 trillion tokens |
| V4 Flash | About 284 billion total parameters, about 13 billion active | Lighter and faster, easier to self host |
How It Benchmarks Against GPT, Claude, and Gemini
DeepSeek V4 Pro’s coding and math scores are genuinely competitive with closed models, not just “good for an open model.” It reportedly scores 80.6% on SWE-bench Verified, ahead of GPT-5.4’s 78.2% on the same benchmark, and matches or exceeds Claude Opus 4.6 on several coding tests. On LiveCodeBench it reaches roughly 93.5%, and on competitive programming it scores around 3206 on Codeforces, a rating that would place it 23rd among all human competitors on that platform. Math benchmarks are similarly strong, with reported scores around 94.3% on AIME 2026 and 95.2% on HMMT. The trade-off shows up on broader general knowledge benchmarks like MMLU-Pro, where it trails the top closed models, so it reads as a coding-and-math specialist rather than an all-around leader.
Key Features
- 1 million token context window by default
- Open weights under the MIT license, available on Hugging Face
- Separate Thinking and Non-Thinking reasoning modes
- Compatible with OpenAI ChatCompletions and Anthropic API formats
- Strong results on reasoning, coding, and agent tasks
Pricing
DeepSeek changed its pricing structure on August 16, 2026, introducing peak and off-peak rates rather than a single flat price. Off-peak V4 Pro API pricing runs about $0.66 per million input tokens and $1.98 per million output tokens, while peak-hour pricing roughly doubles both rates. Peak windows fall between 01:00 to 04:00 and 06:00 to 10:00 UTC, so scheduling large batch jobs outside those windows can meaningfully cut costs. Prices change, so always check the official DeepSeek pricing page before budgeting a project around them.
Limits
V4 Pro still trails the very best closed models on some of the hardest reasoning and general knowledge tests, even where it leads on coding-specific benchmarks. Running the models yourself needs large hardware, especially for the 1.6 trillion parameter Pro model; realistically, self-hosting the full Pro model requires a multi-GPU server setup well beyond consumer hardware, which is why most users access it through the hosted API or the Flash version instead.
Privacy and Use
Check where your data is processed and your company rules before sending sensitive information to any hosted model, including DeepSeek’s official API, since data handling and jurisdiction can differ meaningfully from US-based providers. Self-hosting the open weights avoids that concern entirely but requires the hardware described above.
Read also OpenAI vs Anthropic
Read also AI vs LLM
How Big Has DeepSeek Actually Gotten?
DeepSeek’s rise has cooled from its initial shock but settled into real, sustained usage rather than fading away. Global AI chatbot web traffic share peaked at 12.1% in February 2025 right after the R1 release, then normalized down to about 4.1% by May 2026, still a meaningful share compared to ChatGPT’s roughly 53.9%. In the US specifically, DeepSeek’s share sits around just 1.2%, constrained by government and enterprise restrictions on Chinese AI providers. Usage looks very different inside China, where DeepSeek reported 143 million monthly active users as of August 2025, second only to ByteDance’s Doubao, and the app crossed 23 million downloads within just 19 days of launch, faster than ChatGPT’s equivalent early trajectory. On the business side, DeepSeek says its APIs had been adopted by more than 26,000 enterprises by the end of 2025.
The training cost claims that started this whole story are narrower than headlines suggested. DeepSeek’s own technical report put the final training run for DeepSeek-V3 at 2.788 million H800 GPU hours, roughly $5.6 million in compute at $2 per GPU hour, but that figure covers only the final training run and excludes research, failed experiments, personnel, and infrastructure costs that any lab incurs. It was still dramatically cheaper than the widely reported budgets behind comparable closed models, which is what triggered the initial sell-off in Nvidia and other AI infrastructure stocks in early 2025.
Frequently Asked Questions
What is DeepSeek V4?
A family of open weight language models from DeepSeek, released in Pro and Flash versions on April 24, 2026, both supporting a 1 million token context window.
Is DeepSeek V4 open source?
Yes, the weights are released under the MIT license and available on Hugging Face, which allows unrestricted commercial use and modification, unlike closed models such as GPT or Claude.
Is it cheap?
Yes, relative to closed frontier models. Off-peak API pricing is roughly $0.66 per million input tokens and $1.98 per million output tokens as of August 2026, though peak-hour pricing roughly doubles those rates. Check current rates before budgeting.
Can I run it locally?
Only with very large hardware for the full 1.6 trillion parameter Pro model, typically a multi-GPU server setup. Smaller setups may use the 284 billion parameter Flash model instead, which is considerably lighter to self host.
How does it compare to GPT-5.4 and Claude Opus?
On coding benchmarks specifically, V4 Pro reportedly edges out GPT-5.4 on SWE-bench Verified (80.6% vs 78.2%) and matches or exceeds Claude Opus 4.6 on several coding tests. It trails on broader general knowledge benchmarks like MMLU-Pro, so it is best thought of as a coding and math specialist rather than a universal replacement.
How popular is DeepSeek compared to ChatGPT?
Much smaller globally, but far from negligible. DeepSeek held roughly 4.1% of global AI chatbot web traffic in May 2026 compared to ChatGPT’s approximately 53.9%, though DeepSeek’s usage inside China is much larger, with 143 million monthly active users reported there.
Was DeepSeek really trained for only a few million dollars?
The often-cited figure, around $5.6 million, comes from DeepSeek’s own technical report and covers only the final training run for DeepSeek-V3, not total research, personnel, or infrastructure costs. It was still notably cheaper than typical closed-model training budgets, which is why it shook markets when it was first published.
Why did DeepSeek’s rise affect Nvidia’s stock price?
DeepSeek’s reported training efficiency suggested frontier-level AI models might need far less specialized compute than markets had priced in, which briefly rattled Nvidia and other AI infrastructure stocks in early 2025. The longer-term picture has been more nuanced, since overall AI compute demand has continued growing even as efficiency techniques like DeepSeek’s have spread across the industry.
Final Verdict
DeepSeek V4 is a serious open model with strong value, especially for coding and math work, and its open MIT license makes it genuinely usable for commercial projects without licensing friction. Test it on your own tasks, factor in the new peak and off-peak pricing if you are calling the API at scale, and review data handling before you rely on it for sensitive work.
