AWS has published a benchmarking harness that compares OpenAI models by the cost of a correct answer or usable deliverable. Its September 11 study tests GPT-5.6 Luna, Terra and Sol on Amazon Bedrock against GPT-5.4 Mini and Nano on the OpenAI API.
Across 50 DeepSearchQA web-research questions, Mini averaged 7.6 turns and sent 114,000 input tokens per question, compared with Terra's 50,000. Each turn included the growing conversation history. Luna recorded a cost of $0.05 per passing answer against Mini's $0.40 under the study's pricing assumptions.
For document work, the harness graded 48 GDPval tasks against professional rubrics. Luna passed 27 tasks and Mini passed 20. An 8,192-token output cap truncated some responses, so those results include the consequences of that limit.
The configurations were not equivalent: the Bedrock models ran with reasoning disabled, while the OpenAI API baselines used their defaults. AWS describes a deployment comparison, not an isolated test of model capability. The published harness lets teams replace these samples with their own tasks and measure cost alongside their acceptance criteria.