DeepSeek V4 Pro: Open-Source Frontier, 1M Context, 50–100× Cheaper
DeepSeek V4 Pro launched this week as a fully open-weight, MIT-licensed model with a 1-million-token context window and list pricing 50–100× below premium closed-source rivals. NIST's Center for AI Safety and Innovation (CAISI) confirmed it as the most capable Chinese AI model evaluated to date — while placing it approximately eight months behind leading US models on capability benchmarks. Four independent intelligence sources corroborated the story within the same 48-hour window.
What the Source Actually Says
The DeepSeek TUI repository, which targets V4 Pro as its primary backend, publishes the official pricing table: $0.435/M input tokens (cache miss) and $0.87/M output at a 75% promotional discount valid until 2026-05-05 15:59 UTC, with base list rates around $1.74/$3.48 per million tokens. For comparison, Anthropic Opus lists at $5/$25 per million — making V4 Pro roughly 6× cheaper at standard rates. AI Advances framed the advantage more aggressively on practical workloads: 50–100× lower cost than premium proprietary models on code, agent orchestration, and long-document tasks, backed by V4 Pro's MIT license and hybrid attention architecture.
The model is a 1.6-trillion-parameter Mixture-of-Experts built for Huawei hardware and accessible through DeepSeek's own API, NVIDIA NIM, Fireworks, and self-hosted SGLang. Lev Selector's weekly AI digest classified it as the new open-source frontier ceiling — part of a structural Chinese open-source sweep now comprising seven near-trillion-parameter MoE models, most open-weight, versus a US frontier that remains almost entirely closed. The MIT license removes commercial-use ambiguity entirely.
NIST's CAISI evaluation provides the capability anchor: V4 Pro leads all Chinese models benchmarked while trailing US leaders by approximately eight months. DeepSeek's own documentation, surfaced in the AI Advances analysis, acknowledges a 3–6 month gap on general reasoning — framing it as largely irrelevant for teams whose primary question is whether a competitive model can sustain production agent workflows at scale.
Strategic Take
For teams running cost-sensitive agentic pipelines — code review, document synthesis, parallel sub-agent orchestration — V4 Pro's MIT license, 1M context, and sub-$1/M pricing warrants immediate evaluation, especially while the launch-window discount runs. The NIST 8-month gap matters for frontier reasoning; for the bulk of verifiable production work, the cost curve is the deciding variable.
