DeepSeek Drops V4 Pro and V4 Flash — One Year Later, China's AI Upstart Still Has Something to Prove
DeepSeek launched official preview versions of V4-Pro and V4-Flash on April 24, 2026 — exactly one year after the original DeepSeek shock roiled global markets. V4-Pro targets demanding tasks with frontier-level performance; V4-Flash undercuts GPT-5.5, Gemini 3.1 Pro, and Claude Opus 4.7 on price.
Original sourceOne year ago, DeepSeek rattled the AI establishment by releasing a model that matched U.S. frontier capabilities at a fraction of the training cost. Today, the Hangzhou-based startup is back with V4-Pro and V4-Flash — a two-pronged release that once again puts pressure on OpenAI, Google, and Anthropic to justify their pricing.
## What's New
DeepSeek V4 introduces a Hybrid Attention Architecture that the company says dramatically improves long-context retention — critical for agentic tasks where the model needs to remember queries across extended workflows. Both variants support a 1 million-token context window, enabling entire codebases or lengthy documents to be processed as a single prompt.
Performance improvements are concentrated in coding and math benchmarks, where DeepSeek claims V4-Pro has nearly closed the gap with current frontier models. The company describes the architecture improvements as "drastic" reductions in computational cost compared to V3.2.
## The Price Play
V4-Flash: **$0.14/M input tokens, $0.28/M output** — undercutting GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5.
V4-Pro: **$0.145/M input, $3.48/M output** — undercutting Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4. This is DeepSeek's signature move: competitive capability at commodity pricing.
## Market Reaction
Markets and the AI community are watching carefully but with more skepticism than last January. DeepSeek's original shock was unexpected; this time, everyone knew V4 was coming. The real test will be independent benchmarks from researchers outside China — and whether the open-source Apache 2.0 release holds up to scrutiny on both performance claims and training methodology.
## What It Means
DeepSeek's persistence in the cost-efficiency game is doing something structurally important: it's forcing Western AI companies to compete on price, not just capability. If V4-Pro's claims hold up, it puts another ceiling on what premium models can charge. The AI pricing wars that started with DeepSeek in 2025 are nowhere near over.
Panel Takes
The Builder
Developer Perspective
“V4-Flash at $0.14/M input is genuinely competitive with the best cheap models from OpenAI and Google. If performance claims on coding benchmarks hold up in third-party testing, this is my new default for cost-sensitive agentic pipelines.”
The Skeptic
Reality Check
“The original DeepSeek surprise worked because nobody saw it coming. V4 is expected — Wall Street barely moved. We need independent benchmark verification before trusting the 'closed the gap' claim, especially on reasoning and agentic tasks where the devil is in the details.”
The Futurist
Big Picture
“DeepSeek is systematically dismantling the pricing power of Western AI companies. Hybrid Attention Architecture at scale + 1M context + Apache 2.0 is a template for how sovereign AI ecosystems can compete globally. The AI cost floor just dropped again.”