DeepSeek V4-Pro Lands at $1.74 / $3.48 per Million Tokens — 98% Cheaper Than GPT-5.5 Pro
DeepSeek released V4-Pro and V4-Flash hours after OpenAI's GPT-5.5 announcement, pricing V4-Pro at $1.74 input / $3.48 output per million tokens — roughly 98% below GPT-5.5 Pro's $30 / $180 — under an MIT-licensed open-weight release.
DeepSeek released V4-Pro and V4-Flash on April 24, hours after OpenAI announced GPT-5.5. Both ship under an MIT open-weight license with a 1 million-token context window, undercutting frontier closed-source pricing by an order of magnitude.
The pricing line
V4-Pro is listed at $1.74 input / $3.48 output per million tokens; GPT-5.5 Pro is priced at $30 / $180. That puts the Pro tier at roughly 98% below OpenAI's per-token rate. V4-Flash is cheaper still at $0.14 / $0.28. Architecturally, V4-Pro is a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active per inference; V4-Flash carries 284B / 13B. Both use compressed sparse and heavily compressed attention layers and a new "interleaved thinking" mode aimed at multi-step agent workflows.
Hardware and the export-controls subtext
DeepSeek says portions of V4 were trained on Huawei Ascend chips, sidestepping Nvidia export restrictions in place since 2022. On benchmarks the company shares: V4-Pro-Max scores 3,206 on Codeforces, 80.6% on SWE-Verified (matching Claude Opus 4.6), and 87.5% on MMLU-Pro.
Our Take
The 98% gap is the headline, but the more durable signal is the MIT license at this parameter scale. A 1.6T-parameter open-weight model with 1M context is the first credible self-host alternative for enterprise teams that have been blocked from frontier APIs by data residency or compliance constraints. The Huawei training detail also matters: it suggests the China-side compute supply chain has reached a point where flagship-tier training runs no longer require Nvidia silicon. If DeepSeek's benchmarks hold up under independent evaluation over the coming weeks, the floor on commercial inference pricing has just collapsed.
Want every AI × Web3 signal the moment it breaks? Subscribe to the BlockAI News daily brief.
How we report: This article cites primary sources, regulatory filings, and on-chain data where available. BlockAI News uses AI tools to assist with research and first-draft generation; every article is reviewed and edited by a human editor before publication. Read our full How We Report page, Editorial Policy, AI Use Policy, and Corrections Policy.
