OpenAI has introduced GPT-6 Sol and GPT-6 Luna, two faster, lower-priced models in its GPT-6 family. Its announcement positions Sol for more demanding professional and coding tasks and Luna for lower-cost work at scale. Both are available through the API, with a staged rollout across OpenAI’s work products.
The practical answer for Philippine teams: The lower published token prices make it easier to test AI in support, research, software development and internal workflows. They do not establish that a particular workflow is accurate, safe with personal data or cheaper overall. Test your own tasks and count the cost of tools, retries, review and errors before switching.
What changed in pricing?
OpenAI lists the following API prices per one million input or output tokens, in US dollars:
| Model | Input | Output |
|---|---|---|
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
Its comparison table lists GPT-5.6 Sol at $4 input and $20 output, and GPT-5.6 Luna at $0.20 input and $1.20 output. Thus Sol’s listed input and output rates are each half the earlier listed rates; Luna’s listed input rate is half and its output rate drops from $1.20 to $0.50. These are API token rates, not a guarantee that every application bill will fall by the same percentage. OpenAI also says cached input-token reads can receive a 90% discount under its GPT-6 caching improvements. Check the current official announcement before budgeting.
How do Sol and Luna compare?
OpenAI reports that Sol performed better than GPT-5.6 Sol on its factuality evaluation and made about half as many mistakes in the tested conversations. Those conversations were selected because users had flagged previous model errors; they are not a representative sample of routine use. OpenAI also reports gains for both models in coding, computer use and professional-work benchmarks. Scores depend on the test, reasoning effort, tools and cost calculation. They should be read as vendor-reported benchmark results, not as a prediction for a Philippine law office, BPO or online store.
An initial routing approach is straightforward:
| Task | Model to test first | Review before use |
|---|---|---|
| Short classification, triage and high-volume drafts | Luna | Missed exceptions, language accuracy and escalation |
| Multi-step research, coding and complex document work | Sol | Citations, calculations, tool actions and completeness |
| Decisions with legal, financial, employment or safety effects | Compare both with human review | Error severity, source quality and accountable approval |
This is a testing strategy, not an OpenAI product requirement. A simpler model may handle a narrow task well; a stronger model can still make a consequential mistake.
What should Philippine businesses test?
- Use real task examples with permission. Include Filipino, English, Taglish and the regional languages your users actually use. A general benchmark is no substitute for a Cebuano customer message or a local regulatory question.
- Measure complete workflow cost. Record input and output tokens, cache hits, tool calls, failed attempts, staff review time and any rework after an incorrect answer. A cheap token can support an expensive process if the system repeatedly needs correction.
- Check sensitive inputs and vendor settings. Map the personal and confidential data entering prompts, files, connectors and logs. Review account controls, vendor terms and retention before processing customer or employee information. CyberCode’s AI and data privacy guide explains the Philippine privacy questions; its customer-data-to-ChatGPT guide addresses a common upload scenario.
- Keep a human approval point for consequential output. Verify legal authorities, quotations, amounts and proposed actions against original sources. Give users a clear way to correct or appeal an AI-assisted result. See CyberCode’s AI law and governance hub for the broader Philippine framework.
- Retest after changes. Keep a small fixed evaluation set and rerun it when the model, prompt, tools or knowledge source changes. Record both accuracy and the kinds of errors that matter to your users.
Availability and limits
OpenAI says Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, subject to a gradual rollout. It says Free and Go users can access Luna in the desktop app, while the models are not yet available in Chat. In the API, their names are gpt-6-sol and gpt-6-luna. Availability can change, and a user’s account or region may show the models at a different point during rollout; check the product interface and current OpenAI documentation.
The release changes model options and price points. It does not change Philippine privacy, consumer, employment, copyright or professional duties. For organizations handling personal information, identify the purpose, data flows, vendor role and human decision process before deployment. The best model choice is the one that meets the task’s accuracy and control requirements at an acceptable full-workflow cost.
Primary source: OpenAI, “Introducing GPT-6 Sol and Luna”, accessed September 24, 2026.