Luna’s AI benchmarking cluster: the complete comparison archive

GPT-6 Luna produced this complete cluster during our 22 October 2026 editorial experiment. It received 85.93/100 in the frozen model review and finished production in 9 minutes 28 seconds. Sol’s cluster received 92.05/100 and is the featured blog series.

These are publication copies with disclosed corrections. Scores apply to the original submission. The judge was one GPT-6.1 Sol session, the same model designation as the winning producer; this was not independent human validation.

Read the complete series

• How to benchmark AI models for your own workload — master guide
• Choose metrics and rubrics that reward the right AI behaviour
• Evaluate the agent system, not only its language model
• How to compare AI models fairly
• How to design an AI evaluation around a real task
• How to read public AI model benchmarks without overreading them
• Keep AI evaluations useful as models and products change
• Keep your AI benchmark honest: datasets, holdouts and leakage
• Measure AI latency and cost per successful task
• Repeatability and uncertainty in AI model evaluations
• When humans and model judges should score AI outputs

Compare and verify

• Results and limitations
• Methodology
• Download Luna’s scored original
• Publication edit log
• Experiment evidence and Sol’s featured series

The archive preserves the runner-up’s complete contribution. The publication corrections do not change its frozen score.