DeepSeek-V4 arrives with near state-of-the-art intelligence at 1/6th the cost of Opus 4.7, GPT-5.5

2026-04-24 · Source: VentureBeat · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation, Software Development & Engineering · Depth: Advanced, long

Summary

DeepSeek, a Chinese AI startup, has released DeepSeek-V4, a 1.6-trillion-parameter Mixture-of-Experts (MoE) model under the permissive MIT License. Launched on April 24, 2026, this model offers near-frontier performance, on some benchmarks surpassing advanced closed-source systems, at approximately one-sixth the API cost. DeepSeek-V4-Pro is priced at $1.74 per million input tokens and $3.48 per million output tokens, significantly undercutting OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7. The model features a native one-million-token context window, achieved through a Hybrid Attention Architecture and Manifold-Constrained Hyper-Connections (mHC), reducing KV cache requirements by 90%. DeepSeek-V4 also introduces a two-stage training paradigm and three "effort" reasoning modes, and has been validated on Huawei Ascend NPUs, demonstrating a 1.50x to 1.73x speedup on non-Nvidia platforms.

Key takeaway

For AI Architects evaluating large language models for enterprise deployment, DeepSeek-V4's combination of near-frontier performance and significantly lower API pricing (1/6th to 1/10th of competitors) necessitates a re-evaluation of your cost-benefit calculations. Its open-source MIT license and validated performance on non-Nvidia hardware also offer strategic advantages for supply chain resilience and customizability, making it a compelling alternative to premium closed models for automating tasks previously deemed too expensive.

Key insights

DeepSeek-V4 offers frontier-class AI performance at dramatically reduced costs through architectural innovation and open-source licensing.

Principles

• Architectural innovation can reduce compute costs.
• Open-source models can challenge proprietary leaders.
• Hybrid attention improves long-context efficiency.

Method

DeepSeek-V4 uses a two-stage training: Independent Expert Cultivation via SFT/RL with GRPO, followed by Unified Model Consolidation using on-policy distillation to integrate specialized skills into a cohesive model.

In practice

• Utilize DeepSeek-V4 for cost-sensitive, large inference workloads.
• Explore DeepSeek-V4-Flash for extreme cost efficiency.
• Leverage the 1M token context for complex agentic tasks.

Topics

• DeepSeek-V4
• Mixture-of-Experts
• AI Model Pricing
• Benchmark Performance
• Manifold-Constrained Hyper-Connections

Best for: CTO, AI Architect, Investor, AI Engineer, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• DeepSeek V4 - almost on the frontier, a fraction of the price — DeepSeek V4 models offer competitive performance and industry-leading low costs through significant efficiency gains.
• My Honest Thoughts about Deepseek — DeepSeek V4 offers frontier-level AI capabilities at a fraction of the cost, challenging US AI dominance.
• Deepseek V4 May Disrupt The Entire AI Economy — Deep Seek V4 offers near-SOTA performance at significantly lower costs and provides open-source deployment flexibility.
• As agentic AI pushes rivals to raise prices and cap usage, Deepseek ships a good-enough model for almost nothing — Deepseek's V4 models offer massive context windows and competitive performance at significantly reduced costs via architectural innovation.
• DeepSeek resurfaces with cheap, capable V4 — DeepSeek V4 offers competitive AI model performance at lower costs, supported by Huawei chips.
• Three reasons why DeepSeek’s new model matters — DeepSeek V4 offers competitive performance and cost-efficiency with a 1-million-token context window, optimized for Chinese chips.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by VentureBeat.