Gpt 6 Sol And Luna Push The Cost Efficiency Frontier

An unchanged aggregate score can conceal changes that matter to a workflow. Artificial Analysis' GPT-6 Sol and Luna comparison reported much lower task costs alongside a mixture of gains and regressions. Some professional-work deliverables lost ground because they omitted requirements or had weaker presentation.

The OpenAI models also showed fewer unsupported answers on a knowledge test, partly through declining to answer more often. That can improve reliability while reducing coverage. The release is therefore best read as several tradeoffs: cost, completeness, presentation and willingness to abstain may move in different directions even when a single capability index looks stable.