The price cut is a starting question
OpenAI’s September 22 announcement added GPT-6 Sol and Luna to its model family, positioning them around different balances of capability and cost. The provider also reported evaluation results. We are treating those results as attributed claims, not as measurements of Silver Owl’s work.
The appealing idea is straightforward: if capable inference becomes more affordable, more tasks become worth trying. The harder question is what counts as the task. Is it the model producing a draft, or the business receiving something correct enough to use?
Source: Introducing GPT-6 Sol and Luna · 2026-09-22 · Primary source context · attributed account
Source: Expanding OpenAI Academy with new learning paths · 2026-09-21 · Primary source context · attributed account
Price the complete attempt
Our opinion is that the unit of comparison should be accepted work. Token charges form one part of that cost. Review, retries, missing context and integration can form others. A lower model price is real information; it does not tell us which of those other costs will move.
Take a hypothetical analyst drafting a supplier comparison. One model might produce a fluent answer quickly but leave several claims requiring reconstruction. Another might take longer yet supply verifiable references. The cheaper first response need not be the cheaper finished comparison. Equally, a stronger model may waste money on a task whose requirements are simple and stable.
A meaningful trial would hold the task, permitted sources and acceptance criteria steady. Record the model version and settings. Inspect failed attempts as well as successes. Count the human effort needed to reach a usable result. This is a proposed evaluation method, not a report of a trial we have conducted.
A benchmark is not a verdict on superintelligence
SI means superintelligence in this Journal. A vendor’s coding or workflow evaluation should not silently become evidence that superintelligence has arrived. That is a much broader claim, requiring a clear definition and evidence that survives changes in task, environment and available tools.
One useful discipline is to separate capability, reliability and authority. A system may solve a difficult problem, fail unpredictably on a simpler one, and still have no permission to take consequential action. Collapsing those three questions into a single adjective makes both purchasing and safety discussions worse.
OpenAI’s September 21 Academy expansion offers another reminder that deployment includes people: it describes role-based practice and assessments. That announcement does not independently establish workplace competence. Our view is that training should help staff recognize an unsupported result and know when to stop.
Cheaper intelligence is worth investigating. The purchase decision should follow a bounded comparison of finished work, with the costs of correction visible. Until then, the price list is an invitation to experiment, and the superintelligence argument remains a separate conversation.