What actually changed on September 22
OpenAI introduced GPT-6 Sol and GPT-6 Luna on September 22, 2026. Both are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with a gradual rollout; Free and Go users can access Luna in the desktop app. The company says the models are not yet available in Chat, so product surface and account rollout matter before you redesign a workflow around them.
For API users, the published prices are $2 per million input tokens and $10 per million output tokens for Sol, versus $0.10 input and $0.50 output for Luna. That is a 20× unit-price gap in both directions. It is not a promise of a 20× cheaper completed task: long outputs, retries, tool calls, cache use and human review can change the real cost.
The Method & Miles routing rule
Choose with two questions. First, how costly is a wrong answer? Second, how easy is the output to verify? When both risk and review cost are low, Luna is the economical default. When either rises, Sol is the safer starting point. If the job is unusually consequential and quality matters more than price or speed, OpenAI recommends Astra for its strongest results.
| Task shape | Default | Why | Human gate |
|---|
| Reversible and easy to inspect | Luna | Lower unit cost leaves room for iteration | Quick factual check |
| Multi-source, ambiguous or difficult to inspect | Sol | More capability can reduce hidden rework | Source-by-source review |
| High consequence or uncompromising quality | Astra or specialist review | Do not optimize only for model price | Named accountable approver |
1. Summarizing an inbox or a long document: start with Luna
A summary is usually reversible, and you can compare it with the source. Give Luna the original material, ask for a fixed structure, and require a separate list of uncertain points. Before using the result, spot-check names, dates, numbers and anything framed as a commitment.
Switch to Sol when the summary must reconcile conflicting documents, retain subtle exceptions or support an external decision. The cost of a missed clause can exceed the saving from a cheaper first pass.
2. Drafting an internal email or status update: Luna first
For a low-stakes internal draft, Luna is the sensible default. Supply the audience, desired action, facts that must appear and a word limit. You should still read the final message, but tone and wording are easy to correct.
Use Sol when the message handles a dispute, explains a policy change, combines several stakeholders' positions or will be forwarded outside the team. The model choice should follow the consequence of misinterpretation, not the length of the email.
3. Cleaning meeting notes and assigning actions: Luna with a source check
Luna can turn notes into decisions, owners and deadlines efficiently. Keep the transcript or original notes beside the output, then verify every owner and date. Never let a model invent agreement where the meeting record is ambiguous.
Escalate to Sol when the discussion spans several meetings, contains competing decisions or must be reconciled with a project plan. In that case, require a traceable citation to the source passage for every action.
4. Building a multi-source research brief: prefer Sol
This job is difficult to verify because a fluent brief can hide an omitted source, a stale claim or a false connection. Sol is the better default when the work requires browsing, comparing several documents or preserving caveats. Ask for a claim-to-source table and mark every inference separately from reported fact.
OpenAI reports that Sol makes about half as many mistakes as its predecessor on an internal evaluation built from conversations where users flagged factual errors. That does not establish an error rate for your work, and the evaluation is not representative of typical use. Treat it as directional evidence, not permission to skip verification.
5. Reviewing a spreadsheet or business document: choose by auditability
Luna is suitable for finding obvious formatting issues, extracting a defined set of fields or generating a first list of questions from non-sensitive sample data. Sol is a better starting point for cross-sheet reasoning, inconsistent definitions or a recommendation that depends on several calculations.
Neither model should be the final signer for payroll, tax, legal, medical or regulated decisions. Preserve the original file, show formulas and assumptions, and have a qualified person approve any consequential conclusion.
6. Coding, automation and computer-use workflows: Sol for anything that can change state
OpenAI positions both models for coding and computer use, but a workflow that can send a message, edit a record, spend money or change permissions has a different risk profile from a code explanation. Use Luna for small, isolated transformations with tests. Use Sol for longer, multi-step work where missed context would be expensive.
Whichever model you choose, keep destructive or external actions behind explicit approval, run code in a constrained environment, and verify the actual output. OpenAI's system card treats both Sol and Luna as High capability in cybersecurity and biological/chemical domains, while below Critical in the reported assessments; that is a reason to maintain safeguards, not a consumer performance ranking.
7. Customer-facing copy, contracts and public claims: Sol drafts, a person approves
Use Sol when the draft represents your company, depends on several source documents or contains claims that would be costly to retract. Require a checklist for dates, prices, product names, legal claims and links. A named person should own the final release.
Luna can still produce variants after the facts and approved message are locked. This separates expensive reasoning from inexpensive adaptation: Sol establishes the verified core, while Luna creates controlled versions that remain inside it.
Run a seven-day routing test before standardizing
- Choose five recurring tasks, including at least one easy-to-check and one hard-to-check task.
- Record the model, reasoning effort, completion time, retries and human review minutes.
- Count material corrections, not stylistic preferences.
- Calculate completed-task cost for API work; do not compare token prices alone.
- Keep Luna where it passes the quality gate and move only the difficult cases to Sol.
The useful outcome is not one universal winner. It is a small routing policy that sends each task to the least expensive model that can meet an explicit quality and review standard. For a broader evaluation template, see the Method & Miles AI-tool selection checklist ↗.
