Executive Summary
Vercel has published a report analyzing production data from its AI Gateway service, revealing key trends in the AI model market. The findings show a clear market split: Anthropic commands the largest share of customer spend (61%) on high-stakes reasoning tasks, while Google leads in raw token volume (38%) with low-cost models. The report concludes that there is no single market winner; instead, production applications are becoming increasingly sophisticated, using a multi-model, agentic approach to route different tasks to the most suitable provider based on the workload's cost of failure.
Key Takeaways
* Spend vs. Volume Leadership: Anthropic leads the market in spend (61%), driven by high-cost models for critical tasks. Google leads in token volume (38%), powered by cheaper models like Gemini Flash for high-volume, low-stakes use cases.
* Workload Specialization: The market is fragmented by use case. Anthropic over-indexes on software building, Google dominates consumer applications, and OpenAI has the most balanced distribution across workloads.
* Rise of Agentic AI: Agentic applications that use "tool calls" have doubled their token share in six months, now accounting for 59% of all token volume. These requests are more complex and token-heavy than simple chat.
* Multi-Model is Standard at Scale: Large-scale teams are not locked into a single lab. Teams with over 10 million monthly requests use an average of 35 distinct models, enabling them to optimize for cost, performance, and resilience.
* Rapid Adoption of New Models: Production traffic shifts to new, superior model versions within weeks of release, as multi-model architectures make upgrading a simple configuration change.
* Hidden Cost of Outages: About 5% of spending on AI models is for requests that initially failed and were rescued by a fallback to another provider, indicating that the true cost of unreliability is higher than simple uptime metrics suggest.
Strategic Importance
This report reframes the AI race not as a contest for a single best model, but as competition for specific layers of the application stack. It validates the strategic value of model-agnostic platforms like Vercel's AI Gateway, which enable developers to navigate a complex and fragmented provider landscape effectively.