George Cameron is co-founder and chief product officer of Artificial Analysis, the independent AI benchmarking company he founded with Micah Hill-Smith. He helps developers choose models based on the combined realities of capability, inference speed, cost, and application-specific performance.
Before founding Artificial Analysis, Cameron was a senior strategy consultant at Altman Solon, specializing in technology companies and data centers. He now applies that infrastructure perspective to benchmarking models, inference providers, accelerators, and modalities including language, image, video, and speech. The company’s Artificial Analysis Intelligence Index measures general model capability, while its State of AI research tracks model economics, hardware, and enterprise adoption.
What Cameron measures
- Reasoning changes the real cost of intelligence. Running the same evaluation required approximately 7 million output tokens from GPT-4.1, 72 million from high-reasoning o4-mini, and 130 million from Gemini 2.5 Pro. Cameron argues that teams should calculate cost per task, including reasoning-token verbosity, instead of treating advertised token prices as the complete economic picture.
- Agent latency compounds across sequential calls. An agent making 30 requests at 10 seconds each takes five minutes; at one second per request, it takes 30 seconds. Cameron treats agentic inference latency as a constraint on which products and workflows are viable, not simply an interface optimization.
- Open models and faster hardware expand practical choices. Cameron has tracked the narrowing gap between open-weight and proprietary models, alongside improvements in sparse architectures, distillation, inference software, and specialized accelerators. He expects compute demand to keep growing as reasoning models and multistep agents consume the efficiencies gained elsewhere.
- Evaluation should reflect the actual workload. Cameron helped launch Optima, which enables organizations to build custom benchmarks around their own applications. He describes tailored evaluation as difficult to execute well and essential for comparing models against real requirements.
