The disappearing AI middle class
7 hours ago
- #Open Source AI
- #Model Routing Strategies
- #AI Pricing Polarization
- OpenAI increased its flagship model GPT-5.5's price to $5 per million input tokens and $30 per million output tokens, positioning it as a premium, integrated product.
- DeepSeek released V4-Pro and V4-Flash with significantly lower prices (e.g., V4-Flash at $0.14/$0.28 per million tokens) under an open-source MIT license, aiming to make AI a commodity infrastructure.
- The AI market is polarizing into two clusters: high-cost, closed products (like OpenAI and Anthropic) and low-cost, open-weight models, thinning the middle tier that developers previously relied on.
- Developers must now choose between paying for integrated solutions or using open infrastructure, leading to more model-agnostic agent harnesses and routing logic across different economic models.
- DeepSeek's V4 models leverage advanced architecture (Mixture of Experts) and hardware support (e.g., Huawei Ascend chips), reducing compute costs and challenging Nvidia's dominance in AI inference.
- Self-hosting becomes more viable with models like V4-Flash, offering cost predictability and control, especially for text-only workloads, while the harness layer adapts to manage routing between economies.