AI Gateway Production Metrics

An analysis of production traffic from 200,000+ teams over the past 7 months by Vercel AI Gateway showed that model competition plays out more clearly in real-world operations than in benchmarks. As of April, spend share was Anthropic 61%, Google 21%, OpenAI 12%, but token usage was the opposite: Google 38%, Anthropic 26%, OpenAI 13%, xAI 10%.

High-volume workloads moved across an average of 30+ models and were not tied to a specific lab. Open-source models are also gradually being adopted, but actual customers mix multiple models by use case rather than relying on a single model.

• Premium reasoning calls concentrated on Claude Opus, while low-cost, high-speed calls concentrated on Gemini Flash.
• OpenAI's spend share tripled in April compared to March following the launch of GPT-5.4/5.5.
• Google's spend share rose from 8% → 21%, driven by the spread of Gemini Flash.

The split between cost and usage was also clear across workloads. Personal assistants accounted for 20% of cost and 40% of tokens, coding agents for 22% of cost and 20% of tokens, back-office agents for 6% of cost and 15% of tokens, and app generation for 7% of cost and 11% of tokens.

• B2C generates volume through low-cost, high-volume calls, while B2B generates spend through more expensive calls.
• Cost per token for B2B is roughly 2x that of B2C.
• Agentic workloads account for 59% of all tokens and have doubled in 6 months.