Which names are documented?
OpenAI describes Sol as a model for complex professional work; the API alias gpt-5.6 points to Sol. Terra is positioned as a balance of capability and cost, while Luna targets cost-sensitive, high-volume work. An OpenAI model named “Soul” could not be confirmed in the official sources checked.
Separate tasks by decision complexity
Clear categories and good examples are particularly important for basic inbox classification. Proposal assessment instead needs to account for conflicting requirements, exceptions and the consequences of a wrong assignment. These tasks need different quality criteria even when they use the same documents.
Our recommendation is to separate routine preparation from consequential decisions. One model can structure the input; borderline cases move to a more demanding check or the responsible person. Validate escalation using rules and sample review, rather than the confidence of the answer.
A selection test for sales
Build a small set of anonymized inquiries: a complete request, one with a conflicting budget, one without a clear owner and one outside your offering. Ask each model to extract the same fields and propose a next action. Sales independently assesses whether that action is correct and actually executable.
- Complete required fields or explicitly mark them unknown.
- Do not invent commitments on price, timing or scope.
- Measure incorrect routing and manual corrections as separate costs.
Total cost rather than token price
The official selection method prioritizes quality, followed by cost and speed. Business costs also include integration, retries, monitoring and rework. Compare cost per accepted result: a result counts only after it meets the requirements defined beforehand.
Check current provider prices and plan limits when making the purchasing decision. API prices do not describe a ChatGPT subscription. A visible model name on one surface also does not establish that the same tools are enabled on another.
Start with an evidence-based default
For each workflow, record the default model, permitted data, acceptance criteria and escalation route. Start by collecting the few errors that genuinely interrupt work and improve the process around them. Frequent model changes make comparison difficult when prompts and sources change at the same time.
The AI System Check helps identify the right task category. A subsequent audit turns it into a bounded pilot with a measurable decision instead of a universal model ranking.
Keep it verifiable
Primary sources
- GPT-5.6 Sol modelSource checked:
- OpenAI model catalogSource checked:
- OpenAI model selectionSource checked:
- Workspace model availabilitySource checked:
