DeepSeek V4 drops with 1M context, aggressive pricing

DeepSeek V4 is a price-to-performance wrecking ball that makes token context essentially free for high-volume developers.

• –The 1M context window utilizes a new Engram memory architecture that offloads static knowledge, keeping inference fast and accurate.
• –Pricing is aggressively low: V4-Flash costs just $0.14/1M input tokens, undercutting competitors by orders of magnitude.
• –With 1 trillion total parameters and 32B active per token, it maintains efficiency without sacrificing multimodal capability.
• –Targeted coding benchmarks show V4 reaching 85% on SWE-bench, positioning it as a primary threat to Claude and GPT-5 for software engineering.
• –A 90% discount on cached tokens ($0.028/1M) incentivizes massive RAG and many-shot prompting workflows.