Most enterprise workloads don’t stay on the frontier anyway. Extraction, summarization, classification, doc comparability, and customer-service help usually work completely properly with smaller, cheaper fashions. OpenAI’s personal pitch for the trio of GPT-5.6 fashions isn’t merely that Sol is best. It’s that Terra and Luna ship totally different mixtures of intelligence, latency, and value. Luna, the most affordable tier, almost matches the earlier era’s peak efficiency at lower than half the estimated value, based on OpenAI.
The sensible query, in fact, is the place to begin. An enterprise can’t check each mannequin, each reasoning setting, and each value tier earlier than doing any work. So right here’s my recommendation (which I don’t comply with in my very own work, however I’m not defining enterprise technique and generally is a little price-insensitive). Start with the most affordable credible mannequin that seems able to the duty. Give it a consultant set of actual examples and, earlier than you begin testing, outline what counts as adequate. If it passes, cease. If it fails, transfer up a tier or strive a mannequin with strengths higher suited to the work.
That sounds virtually offensively easy, however it reverses the best way many individuals, together with me, use these merchandise. We begin with the largest mannequin as a result of we’re afraid of what we would lose. Enterprises ought to begin decrease and require proof earlier than paying for extra intelligence.
