Shift in LLM Traffic Patterns

Data released via a report on Dirac indicates a significant shift in the distribution of Large Language Model (LLM) traffic on the OpenRouter platform. According to unverified reports circulating on Reddit, the DeepSeek model accounted for 53% of all tokens processed on OpenRouter yesterday, totaling approximately 27 trillion tokens.

The data suggests a broader trend toward open-weights models, with total market share for open-weights labs reaching over 75% of the platform’s traffic. While these figures are being discussed within developer communities, they have not been independently confirmed by OpenRouter or official laboratory statements.

Economic Implications of Model Selection

The surge in DeepSeek usage highlights a growing divergence between token volume and revenue share. Analysts noted in community discussions that token share does not necessarily correlate directly with revenue share, particularly as open-weights models often provide a more cost-effective alternative to proprietary models.

Individual user reports on Reddit underscore the cost disparity between different model providers. One developer noted that an intense, agent-driven coding session using DeepSeek cost approximately $4 USD, whereas a similar session using Anthropic’s Claude model reportedly exceeded $400 USD. This significant price gap appears to be a primary driver for the high volume of tokens being routed through open-weights providers.

Strategic Utility of Aggregation Platforms

Despite the availability of direct APIs from individual model developers, OpenRouter continues to see high engagement. Users have identified several strategic reasons for utilizing an aggregator rather than direct integration with providers like DeepSeek, OpenAI, or Anthropic.

Key drivers for using OpenRouter include:

  • Unified Management: The ability to manage a single API key to access a wide variety of models.
  • Provider Redundancy: The ease of swapping between different providers to ensure uptime or optimize performance.
  • Data Privacy: Some users indicated that OpenRouter’s data retention policies may be preferable to those of individual model developers.
  • Granular Usage: The platform allows users to pay for specific, high-end model calls (such as Claude Opus) on a per-use basis rather than committing to monthly subscriptions.

While the high token volume for DeepSeek is notable, observers have cautioned that OpenRouter represents a specific subset of AI traffic, which may not reflect the broader global market share of these models across all platforms and enterprise integrations.