Following the release of OpenAI’s GPT-5.5, DeepSeek has launched its own models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, just hours later. This timing might seem coincidental but is noteworthy given U.S. export bans on chips impacting Chinese AI development for three years.
DeepSeek from Hangzhou introduced preview versions of these models today. Both are open-weight with a one million token context window, allowing operations comparable to handling the Lord of the Rings Trilogy in size. They remain significantly cheaper than Western counterparts and free for local use.
Previously, DeepSeek’s R1 model in January 2025 caused Nvidia’s market cap to plummet $600 billion as investors questioned large investments by American companies compared to the efficiency achieved by a small Chinese lab. V4 marks a quieter yet technical shift towards AI building efficiencies.
DeepSeek-V4-Pro, with its 1.6 trillion parameters and only 49 billion active during inference, represents the largest open-source model in the LLM market to date. Its Mixture-of-Experts technique ensures full knowledge access without increased compute costs for each request. DeepSeek claims V4-Pro-Max as the best open-source model, excelling in coding benchmarks and narrowing performance gaps with closed-source models.
V4-Flash offers 284 billion parameters with 13 billion active. It is faster, more economical, and provides similar reasoning capabilities to Pro when given a larger budget for thinking.
Both support one million tokens of context as a standard feature. DeepSeek openly shares its technical details on GitHub, introducing two new attention types—Compressed Sparse Attention and Heavily Compressed Attention—to address the scaling problem with traditional AI attention mechanisms.
These innovations result in V4-Pro using only 27% of the compute required by its predecessor at one million tokens, while memory needs drop to 10%. V4-Flash reduces these further to 10% of compute and 7% of memory. Consequently, DeepSeek can offer prices significantly lower than competitors: GPT-5.5 Pro costs $30 per million input tokens and $180 for output, compared to DeepSeek V4-Pro’s $1.74 input and $3.48 output, with V4-Flash at $0.14 input and $0.28 output.
Cline CEO Saoud Rizwan highlighted that Uber’s AI budget in 2026 would have lasted seven years using DeepSeek instead of Claude for four months, given these cost reductions.
DeepSeek’s technical report also candidly publishes performance gaps with models like GPT-5.4 and Gemini-3.1-Pro. While V4-Pro trails in some areas like multitasking benchmarks, it excels in codeforces and STEM problem-solving.
Notably, V4-Pro integrates into AI coding tools and outperforms Claude Sonnet on agentic tasks. It ranks first among open-weight models on GDPval-AA, closing the performance gap with Claude Opus 4.6.
The new release also features ‘interleaved thinking,’ maintaining reasoning context across multiple tool calls, enhancing complex automated workflows.
Despite U.S. chip export bans since 2022 aiming to slow Chinese AI progress, DeepSeek has innovated a more efficient architecture and developed domestic hardware capabilities. This launch comes amid significant activity in the AI sector with competitors like Anthropic and Xiaomi also releasing advanced models.
The question now for developers is when premium offerings are worth their price, especially as enterprise-level tasks become economically viable with DeepSeek’s affordable pricing and customization potential through its open-source nature.