How Open Models Win Against OpenAI and Anthropic

Summary

The talk argues that open models have reached a point where they can support useful agentic and enterprise tasks, but their long-term success depends on building a sustainable economy—not just releasing capable models.

• Adoption is growing, but the evidence is incomplete. Public figures from inference providers such as Together AI, Baseten, Fireworks, and OpenRouter suggest rapid growth in open-model usage. Some providers appear to have reached token volumes comparable to Google’s Gemini API. These figures capture only part of the market, since open models are also hosted on major cloud platforms.
• Open models contribute to research and innovation. The speaker’s analysis of arXiv papers found that Qwen is frequently cited in machine-learning research. Open model families can serve as a foundation for further experimentation and development.
• The performance gap is small, but open models are not automatically cheaper. Using Artificial Analysis performance comparisons, the speaker says open models have generally trailed closed models by several months. Closed providers can still offer better performance for a given cost, thanks to advantages in training, reasoning, and compute.
• The open-model economy is fragmented. Model developers bear the high costs of training, while inference providers and companies specializing in post-training or deployment may capture more of the revenue. Overlapping business models make it difficult to see how model developers will reliably fund the next generation of systems.
• Closed labs benefit from vertical integration. Companies such as OpenAI and Anthropic can coordinate products, models, compute, and infrastructure, potentially lowering service costs and directing revenue back into development. Open-model participants lack that unified structure.
• The proposed path to winning is differentiated value. Open models should aim to provide affordable, adaptive, private, and controllable intelligence—especially for organizations that need to use proprietary data, retain control of their models, or specialize systems for particular industries.
• Lower costs and specialization are central. The speaker argues that open-model providers need to push inference prices toward marginal cost and find sustainable revenue through volume, specialist services, and enterprise deals. Customized models may be especially useful for narrow agent tasks where general-purpose systems are unnecessarily costly.
• Early examples point to potential savings. The talk cites work by Harvey and others suggesting that well-trained open models can perform strongly on specific tasks at lower cost. One example claims a Qwen model achieved higher accuracy than leading alternatives on a narrow task.

Reviews, Guides, or Tutorials

None. This is an industry and economic analysis.

Main Speaker and Sources

Nathan Lambert is the speaker. The talk refers to usage data and examples from Together AI, Baseten, Fireworks, OpenRouter, Google, Artificial Analysis, and arXiv, as well as work by Harvey, Merkle, Applied Compute, and Prime Intellect.