# OpenRouter State Of Models: Jev, Open-Weights, and Tokenomics

## Executive summary

The AI landscape is characterized by massive, near-exponential growth in token usage (e.g., 146 trillion tokens last week), indicating strong demand but necessitating a focus on 'tokconomics'—maximizing value per token spent. The market is highly competitive, with open-weights models and specialized classifiers (like Jev) challenging traditional LLM providers. Engineers are advised to move beyond simply using more tokens and instead focus on building complex, multi-model agentic workflows that optimize for performance, speed, and cost.

## Key takeaways

- Token Usage is Skyrocketing: OpenRouter recorded 146 trillion tokens used last week, compared to 4.5 trillion the previous year. This sustained growth suggests strong market demand, arguing against an 'AI bubble' narrative. (03:20)
- Tokconomics is Key to Value: Simply increasing token usage does not equal increased value. Engineers must optimize their workflows to ensure token spend leads to genuinely useful work, rather than just burning tokens for perceived productivity. (04:10)
- Specialized Models are Disrupting LLMs: Zero-shot classifier models, such as Jev, are emerging as a 'new species' of model. They are highly scalable, reliable, and can significantly reduce cost and improve speed when integrated into agents, complementing traditional LLMs. (08:30)
- The Market is Decentralized: The AI race is not won by a single provider. The optimal scenario involves combining compute from multiple sources (e.g., DeepSeek, Gemini, Luna, and Jev) to achieve the best performance/cost trade-off. (07:30)

## Technical details

- OpenRouter Market Data: OpenRouter tracks model usage, showing DeepSeek, Google, and OpenAI consistently in the top three market share spots. The platform is used as a proxy for individual engineer and small-to-medium business spending, contrasting with large enterprise spending. (06:00)
- Model Performance & Trade-offs: Gemini 3.8 Flash is highlighted for its effective balance of performance, speed, and cost, making it a strong workhorse model. (06:30)
- Jev Classifier Model: Jev is a zero-shot classifier model that offers superior scalability and reliability compared to many alternatives. Integrating Jev into agentic workflows can save significant token spend (estimated 20%) and improve speed. (08:30)
- Agentic Engineering: The trend favors owning and customizing agentic coding tools (e.g., PI coding agent) over renting them, providing a significant competitive advantage in workflow control and cost management. (11:00)

## Practical implications

- Focus on building agentic workflows that utilize multiple, specialized models (e.g., LLMs + classifiers) rather than relying on a single, monolithic model.
- Prioritize optimizing for 'tokconomics' by measuring the value derived from token spend, not just the volume.
- Integrate specialized, low-cost models (like Jev) into agents to handle decision-making and classification tasks, thereby reducing overall compute cost and improving speed.
- Develop proprietary agent harnesses to maintain control and customization over the AI workflow, mitigating vendor lock-in.

## Topics

LLMs, Agentic Engineering, Tokenomics, Open-Weights Models, Classifier Models, OpenRouter rankings, Anthropic IPO News, Pi Agent

Source: https://www.youtube.com/watch?v=8BD6w5wELRo