DeepSeek V4-Pro: 1.6 Trillion Parameters, 1M Context Window, and Benchmark Parity With Frontier Models at One-Tenth the Price

DeepSeek released preview access to V4-Pro and V4-Flash on April 24, and the benchmarks are characteristically blunt: parity with GPT-5 and Gemini 3.0 Pro on MMLU-Pro, MATH-500, and HumanEval, at $0.145 per million input tokens — roughly a tenth of what Western frontier labs charge.

V4-Pro is a 1.6-trillion-parameter mixture-of-experts model. The one-million-token context window is genuine, not the cosmetic long-context support that degrades past 128K that some labs ship. Early testers on Hugging Face report clean retrieval of information planted at 800K tokens in multi-document QA tasks. V4-Flash, the lighter sibling, targets latency-sensitive applications with sub-100ms median responses on typical prompts.

The architecture extends DeepSeek-V3’s multi-head latent attention design at considerably larger scale. Training ran on Huawei Ascend clusters — a constraint that has shaped DeepSeek’s efficiency-first approach since US export controls cut off Nvidia H100 access. Inference reportedly sustains 40,000 tokens per second at full load.

On benchmarks: V4-Pro leads GPT-5 marginally on AIME 2025 math reasoning and sits within noise on GPQA Diamond. Long-context needle retrieval edges Gemini 3.0 Pro. Creative writing and nuanced instruction-following remain the gap — Western labs still hold an advantage there.

The pricing reframes the market in concrete terms. Enterprise teams processing large document sets, code repositories, or financial filings will do the math quickly. At $0.145 per million input tokens, a one-billion-token monthly workload costs roughly $145. The same workload on GPT-5 runs ten to fifteen times higher. That difference isn’t theoretical; it’s the number that shows up in quarterly infrastructure budgets.

DeepSeek’s API reliability has historically stumbled under spike traffic. In the first 48 hours of V4 preview, no major outages were reported, suggesting the infrastructure was pre-provisioned at larger scale than previous launches. That matters for production adoption — the capability doesn’t help if the endpoint is unavailable when it’s needed.

The timing is not accidental. DeepSeek is in active fundraising talks at a reported $20 billion valuation, with Alibaba and Tencent both at the table. A credible frontier-competitive model preview weeks before a close shifts negotiating dynamics. V4-Pro is as much a term-sheet move as a technical release.

For developers, the practical question: is V4-Pro a drop-in replacement for Gemini 3.0 Pro Flash on price-sensitive workloads? Early evidence says probably yes, with the caveat that creative and open-ended tasks still favor the Western models. For code generation, retrieval, and structured reasoning — which is most of what production LLM pipelines actually do — V4-Pro is a serious option at a price that makes the alternatives look expensive.

The model is in preview, not GA. Expect stability improvements before the full release, and stress-test context handling above 500K tokens before committing production traffic.