AWS Benchmarks Show GPT‑5.6 Luna Often Beats Mini and Nano on Cost and Quality
AWS compares five OpenAI models on Amazon Bedrock, finding the newer GPT‑5.6 Luna delivers lower cost per correct answer, per agent pass, and per professional deliverable than the older GPT‑5.4 mini and nano.
Original source published: September 11, 2026
AWS released an open‑source benchmarking harness that runs the same OpenAI Responses API against three Bedrock models (gpt‑5.6‑luna, terra, sol) and two cost‑optimized OpenAI API baselines (gpt‑5.4‑mini, nano). Across single‑call accuracy tests (AIME, GPQA Diamond, MMLU‑Pro) Luna achieved the lowest observed cost per correct answer—$0.0021 versus $0.0139 for mini after a July 30, 2026 price reduction. In multi‑turn agent tasks (DeepSearchQA), Luna required fewer turns and cost $0.05 per passing answer, compared with $0.40 for mini, while terra’s higher token price was offset by fewer turns.
For document‑style workloads (GDPval), Luna passed 27 of 48 professional deliverables at $0.010 per pass, beating mini’s $0.030 and nano’s $0.012. The post also notes Luna’s median time‑to‑first‑token was 21 % lower than mini and its throughput 43 % higher for outputs ≥500 tokens. AWS suggests using these outcome‑based metrics—cost per correct answer, turn efficiency, and rubric pass rate—to decide which model fits a given workload.