OpenAI introduced GPT-6 Sol and Luna on 22 September. Against GPT-5.6 promotional rates, Sol’s price per million tokens falls from $4 to $2 for input and from $20 to $10 for output. Luna falls from $0.20 to $0.10 and from $1.20 to $0.50 respectively: the last reduction is 58.3%, calculated from the table. Announcement and prices.
The commercial opportunity is straightforward. When each operation costs less, uses that did not justify integration may start to do so. Classifying documents, preparing replies or comparing records becomes less dependent on the price of a single query. More depends on how work is distributed among differently priced systems, people and automatic checks. The problem begins to resemble organising an office rather than choosing a champion. Different assignments need different capabilities, even when they arrive through the same interface.
The cheaper model may be the better choice
Not every task benefits proportionately from the most expensive model. Extracting a date from a legible document and deciding whether a contractual exception changes an obligation are different operations, although both involve text. If an economical tool performs the first reliably enough for the result to be checked, using more resources may add little. That saving can preserve attention for decisions where a mistaken interpretation changes the outcome rather than merely requiring a field to be corrected.
Imagine a service receiving delivery enquiries. One model identifies an order number and retrieves its status; another process examines disputes over refunds. The organisation has to establish when to pass from one to the other. A customer making spelling mistakes may have a straightforward request. Someone using an apparently routine phrase may be reporting a serious failure. Message length or elegance would be a convenient classification rule, but a poor basis for deciding how much scrutiny a case deserves.
Trouble arises when the system identifying difficult cases is also the one least equipped to understand them. Automatic referral can work well overall yet miss precisely the exceptions that justified its creation. Checking only referred cases leaves these errors out of sight. Some apparently resolved cases need examination too, and recipients need a way to reopen an incorrect classification. Otherwise the apparent success rate partly reflects which mistakes the arrangement allows anybody to discover.
Confidence needs a particular task
In August, NIST released TEVV-Athlon for consultation, a draft framework for constructing evaluations suited to organisational objectives and different applications. It does not compare Sol with Luna or certify either model. Its relevance lies in making explicit what a test needs to measure before selecting it. NIST’s presentation.
A control should address the problem it is meant to solve. In this service that means distinguishing extraction errors, questionable inferences and decisions requiring an authority. Improving the first category does not establish that the others have disappeared. An organisation could economise on document reading and spend some of the saving on situations previously buried among routine assignments. The result would be a different allocation of expertise, rather than simply fewer people looking at every file.
An economic comparison also needs to include round trips. A cheap first attempt followed almost invariably by a second model may cost more than direct referral. Agreement between two systems does not necessarily supply two independent checks: both may share the same mistaken assumption. A persuasive explanation of why a case is simple may itself make subsequent review less likely. The architecture of the service matters as much as the rates charged for its components.
For a small business, lower prices open genuine possibilities: a useful function may no longer require sales volumes sufficient to pay for expensive queries. Integration nevertheless needs an exit. Data, instructions and proprietary tests should remain transferable when prices or capabilities change; otherwise an attractive opening offer purchases dependence paid for later.
The gain would be to use fewer resources for an adequately supported answer while preserving an effective route for cases needing more. Economising by default, with recognisable exceptions, can expand a service. Preventing an exception from reaching someone able to resolve it merely shifts the expense onto somebody else.