DeepSeek Releases V4 Pro, Challenging OpenAI, Anthropic on Key Benchmarks

Chinese AI startup DeepSeek has launched its next-generation flagship model family, DeepSeek-V4. It has introduced two variants, V4 Pro and V4 Flash, as it doubles down on open-source competition in the global AI race.

The launch marks DeepSeek’s first major foundational model release since its breakout R1 model in early 2025. The company said the new V4 series is built for stronger reasoning, coding, and agentic workflows, with both models supporting an ultra-long one-million-token context window.

DeepSeek’s open-weight model is now competitive across coding, reasoning, and agent tasks, while still trailing slightly on peak knowledge benchmarks dominated by closed systems.

On benchmarks, DeepSeek positions V4 Pro Max close to leading closed models from OpenAI and Anthropic, but the gap varies by task. In MMLU-Pro, DeepSeek reports 87.5, broadly in line with top-tier systems, though still behind the highest scores seen from frontier models. 

On coding, it claims a lead in LiveCodeBench with a 93.5 pass rate, edging past several proprietary peers, while also posting a Codeforces rating of 3206, suggesting strong competitive programming ability. In reasoning-heavy math benchmarks such as IMOAnswerBench, DeepSeek narrows the gap but does not consistently surpass the best closed models. 

Where it stands out is long-context evaluation, with stronger performance on 1M-token tasks like MRCR and CorpusQA, and in agentic workflows where results on SWE and terminal-based benchmarks are broadly comparable. The takeaway is not outright dominance, but convergence. 

According to DeepSeek, V4 Pro is the flagship offering with 1.6 trillion parameters, while V4 Flash, with 284 billion parameters, is positioned as a faster and more cost-efficient version for production workloads and API deployments.

In a post on X, the company highlighted that the model has been optimised for AI agents, claiming seamless integration with coding assistants and autonomous developer tools such as Claude Code, OpenClaw, and OpenCode. DeepSeek added that the model is already powering its own internal agentic coding workflows.

The company also announced that the API is live, with support for OpenAI ChatCompletions and Anthropic APIs, allowing developers to switch models by simply updating the model name to deepseek-v4-pro or deepseek-v4-flash. Both versions support thinking and non-thinking modes, a design increasingly common among frontier models focused on reasoning tasks.

Early benchmark claims from the company suggest the V4 Pro model narrows the gap with top-tier proprietary models on coding and reasoning benchmarks, while retaining DeepSeek’s emphasis on cost efficiency.