The era of models is over, we are in the era of harnesses
11 hours ago
- #Market Segmentation
- #Cost Efficiency
- #AI Models
- The AI industry has shifted from focusing on raw models to offering 'harnesses'—systems that route user queries to the most cost-effective AI model based on required intelligence.
- Major AI companies like OpenAI, Google, and Anthropic label their subscription services with flagship model names but often deliver harnessed versions to reduce compute costs and maintain profitability.
- Intelligence scales with compute, which is expensive, leading companies to balance perceived quality with cost efficiency to cater to the consumer price point of around $20/month.
- Subscriptions now provide access to model harnesses rather than direct model access, with companies optimizing the intelligence-to-compute efficiency ratio to make money at consumer price levels.
- Direct access to state-of-the-art (SOTA) models is available via APIs, but users incur higher costs, while smaller, local models are becoming smarter and more affordable, potentially running on personal devices soon.
- The AI market is segmenting into three areas: cheap daily driver models (often local or small-scale), subscription-based harnesses for cost-effective superintelligence, and direct SOTA model access for cutting-edge needs.
- Google is positioned strongly in cheap models due to its large-scale AI integration in services like search, while Anthropic leads in SOTA model access, despite recent concerns about performance degradation and sustainability.
- Users have benefited from subsidized SOTA intelligence in the past and will continue to have access to capable $20/month models, with 'good enough' AI becoming increasingly accessible and affordable.