DeepSeek released V4 on 24 April 2026, exactly one year after its V3 and R1 models had rattled the AI industry. Six months later, it is worth looking back at what V4 actually was, what it changed, and what it means for developers and businesses today.
What DeepSeek V4 is
DeepSeek V4 is a family of two open-weight models built on a mixture-of-experts design.
- V4-Pro has 1.6 trillion total parameters, with about 49 billion activated for each token.
- V4-Flash has 284 billion total parameters, with about 13 billion activated for each token.
Both support up to one million tokens of input and up to 384,000 tokens of output, according to reports on the launch. They offer three reasoning modes, from fast non-thinking replies to maximum-effort reasoning. The weights were published on Hugging Face, and DeepSeek's API was made compatible with both OpenAI's and Anthropic's request formats.
DeepSeek also claimed big efficiency gains. At one million tokens of context, it reported that V4-Pro needed only about 27 percent of the computing per token and 10 percent of the memory cache that its previous model, V3.2, required.
The price shock, again
When DeepSeek-R1 arrived in early 2025, its low cost was the headline. V4 repeated the pattern. Reported API prices were about $1.74 per million input tokens and $3.48 per million output tokens for V4-Pro, and just $0.14 and $0.28 for V4-Flash.
Those prices sit far below many American frontier models. Reports at launch compared V4 to models costing many times more. For developers, this made high-volume AI features dramatically cheaper.
Change one: cheaper AI became the norm
The biggest effect of V4 has been on pricing across the industry. When a capable open model is available at very low cost, other providers face pressure to cut prices or offer cheaper tiers. Through 2026, major American labs released budget models with much lower prices than their flagship systems. OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5 are both examples of very cheap, capable small models.
It is impossible to credit one release for an industry-wide trend, but DeepSeek's repeated low-price launches have clearly shaped expectations. Developers now expect strong performance at a fraction of what it cost a year or two ago.
Change two: open models closed more of the gap
DeepSeek reported self-measured scores of 80.6 percent on SWE-bench Verified, a test of fixing real-world software issues, 93.5 percent on LiveCodeBench, and a Codeforces rating of 3206. These figures came from DeepSeek's own testing and needed independent confirmation, but they placed V4 close to the best closed models on coding.
Together with Moonshot's Kimi K3 in July and Alibaba's Qwen releases, V4 strengthened the case that open-weight models are no longer far behind. For many tasks, the difference between the best open and closed models is now small enough that cost, privacy and control become the deciding factors.
Change three: long context became cheap
One million token context windows used to be a premium feature. V4 offered them at very low prices, and its efficiency claims suggested that long-context processing could be far cheaper to run. This made it practical for developers to build tools that read entire codebases, long legal documents or large document collections without breaking the budget.
Change four: easier switching between providers
By making its API compatible with both OpenAI and Anthropic formats, DeepSeek lowered the cost of switching. Developers could test V4 by changing a few lines of configuration. This reflects a broader trend: AI applications are increasingly built to work with several models, choosing the best one for each task.
What has not changed
Some concerns from the R1 era remain.
Data and privacy. DeepSeek's own app and API process data on servers in China, and its privacy policy allows data storage there. Several governments and organisations have restricted DeepSeek's app on official devices. Running the open weights yourself, or through a trusted provider, avoids sending data to DeepSeek.
Content rules. DeepSeek's hosted services follow Chinese regulations and may avoid or reshape answers on politically sensitive topics.
Hardware demands. V4-Pro is very large. Self-hosting it requires serious infrastructure, so most users rely on the API or hosted providers.
Benchmark verification. As with many launches, headline numbers were self-reported. Independent testing gives a more reliable picture.
What it means for developers in India
For Indian developers and startups, V4 has been particularly significant. Cost is often the biggest barrier to building AI products for a price-sensitive market. V4-Flash, at very low prices, makes features like customer support bots, document processing and multilingual tools affordable at scale. The open weights also offer a path for organisations that need to keep data within their own infrastructure, which matters for sectors such as finance, healthcare and government.
Practical tips:
- Test V4-Flash for high-volume, simpler tasks before assuming you need a more expensive model.
- Use V4-Pro or other strong models for complex coding and reasoning.
- Keep sensitive data off third-party servers unless you are comfortable with their policies.
- Design your application so you can switch models easily as prices and capabilities change.
The bottom line
Six months on, DeepSeek V4's legacy is clear. It pushed AI prices lower, narrowed the gap between open and closed models, made long context affordable and encouraged easy switching between providers. It did not resolve concerns about data privacy and content rules, and its benchmark claims still deserve independent scrutiny. For developers, the lesson is to treat V4 as a serious, low-cost option, test it against your own work, and choose based on results rather than hype.
