Beyond the price per token: Choosing the right OpenAI model on Amazon Bedrock for your workload

This article benchmarks OpenAI models on Amazon Bedrock against cost-efficient OpenAI API baselines, demonstrating that true model selection should prioritize cost per successful outcome rather than token pricing alone.

• Measured cost per correct answer on AIME, GPQA Diamond, and MMLU-Pro benchmarks across five models
• Analyzed agent trajectory costs on DeepSearchQA, showing turn efficiency drives input-token volume quadratically
• Evaluated professional deliverables using GDPval rubric-graded tasks with real occupational standards
• GPT-5.6-Luna on Amazon Bedrock achieved lowest cost per passing outcome across benchmarks after July 2026 repricing
• Provided open-source benchmarking harness for customers to evaluate models on their own workloads
• Decision framework recommends Luna for high-volume tasks, Luna/Terra for agentic workloads, and Sol for accuracy-critical work

Organizations should measure full cost of successes and failures, including token efficiency and turn count, rather than relying on pricing pages alone when selecting models.