GPT-6 Sol and Luna: Pricing, Availability and a Business Test Plan

OpenAI announced GPT-6 Sol and GPT-6 Luna on 22 September 2026 as faster, lower-cost members of the GPT-6 family. The practical news is not simply that two more models exist. Businesses now have a wider cost-and-capability choice for routine automation, coding, analysis and agent work—but the cheapest model on a price card is not automatically the cheapest way to complete a task.

This guide separates the verified release details from the decision a team needs to make. It explains where the models were available at announcement, records the API prices OpenAI published, and gives a bounded test plan for deciding between Luna, Sol and Astra. Product facts were checked against OpenAI's official announcement on 23 September 2026. The worked example below is hypothetical, not a report of a ProdifyDigital test.

What OpenAI announced

The official GPT-6 Sol and Luna announcement positions both models as more cost-efficient ways to use advances introduced with GPT-6 Astra. OpenAI says the family shares improvements in professional work, factuality, coding, computer use, alignment and collaboration style. Astra remains the company's highest-capability option; Sol is aimed at difficult work with more room to iterate, while Luna is the lighter, lower-cost choice.

That description is a useful starting hypothesis, not a universal routing rule. A short classification job, a multi-file coding change and a customer-facing policy analysis have different error costs. The right model is the least expensive option that meets the acceptance standard for the specific workflow, including retries and human review.

GPT-6 Sol and Luna price

OpenAI listed GPT-6 Sol API pricing at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna was listed at $0.10 per million input tokens and $0.50 per million output tokens. The company described both as 50% cheaper than their GPT-5.6 promotional pricing. These were the published prices on 22 September 2026; teams should confirm current pricing and any platform charges before budgeting.

Token rates do not equal completed-task cost. A workflow may also consume cached and uncached context, reasoning, tool calls, retries, generated output and reviewer time. A lower-priced model can cost more overall if it needs repeated correction. A higher-priced model can be wasteful when the work is predictable and easy to verify.

OpenAI also announced improved GPT-6 prompt caching, with eligible cached input-token reads receiving discounts of up to 90%. Its separate prompt-caching announcement explains new dashboards, diagnostics and explicit cache breakpoints. Treat cache savings as workload-dependent: shared prefixes must actually be reusable, and changing tools or instructions can affect the hit rate.

Where the models were available

At announcement, OpenAI said GPT-6 Sol and Luna were available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users could access GPT-6 Luna in the desktop app. The company said the models were not yet available in Chat and that the rollout would continue gradually during the day.

For API users, the announced model identifiers were gpt-6-sol and gpt-6-luna. Availability can still differ by account, surface, region or integration. Check the model picker or API response you actually use rather than assuming a headline applies everywhere.

The date matters. If a reader opens this article later, “available” should mean visible and usable in that reader's account today, not merely announced on 22 September. Record the exact model identifier and test date in any production approval.

How to choose between Luna, Sol and Astra

Use task evidence rather than a single model hierarchy. A practical first routing rule is:

• Start with Luna for high-volume, bounded work with clear inputs, short outputs and inexpensive verification.
• Test Sol when the task needs stronger reasoning, coding or multi-step work but still has a meaningful cost ceiling.
• Reserve Astra for the most demanding or consequential cases where a capability gain is worth the higher cost and review requirement.

This is a test order, not an endorsement of automatic escalation. Some workflows should stop when evidence is missing rather than pass the same prompt to a more powerful model. Actions involving customers, money, security, legal commitments or public publication still need appropriate human authority.

Before choosing a model, use the value, risk and reversibility worksheet to decide whether the workflow is suitable for automation at all. The broader small-business AI control guide covers data boundaries, approval and monitoring.

A six-step business evaluation

Run a small comparison before changing a live workflow. Keep the current method as a baseline and use representative, non-sensitive tasks with observable results.

1. Define one task and owner. State what the model may receive, what it must produce, who reviews it and which actions remain prohibited.
2. Build a small task set. Include normal cases, one edge case and one case with deliberately missing evidence. Remove customer secrets and production credentials.
3. Write acceptance criteria. Score factual support, instruction coverage, unsafe or invented claims, format compliance and reviewer corrections. Avoid “looks good” as a result.
4. Run comparable configurations. Give Luna, Sol and the existing baseline the same evidence, core instructions and tools. Record any reasoning or tool difference rather than hiding it.
5. Measure the whole task. Capture tokens, cache usage, elapsed time, retries, failed tool calls and human review minutes. Separate vendor-listed price from observed workflow cost.
6. Approve a bounded route. Decide which model may handle which class of task, when to escalate, and which failure stops the workflow. Re-test after model, prompt, tool or policy changes.

A hypothetical customer-support evaluation

Imagine a small software company wants AI to draft replies for common account questions. It prepares 20 anonymised cases: password help, billing explanations, plan comparisons, one angry customer, two ambiguous requests and two cases where the correct answer is absent from the approved knowledge base.

The models may draft but not send. Each answer must cite the supplied policy section, avoid promising refunds, and clearly escalate when evidence is missing. A reviewer scores accuracy, policy coverage, unsupported claims, tone, correction time and total API cost.

Luna may be sufficient for straightforward requests. Sol may reduce corrections on ambiguous or multi-part cases. Astra may add little value if the knowledge base already determines the answer—or it may help on the hardest cases. The team should discover that from its own scorecard, not from the model names.

If Luna produces an acceptable answer for 14 cases and safely escalates four, that may be a better route than forcing it to answer all 20. If Sol resolves two of the escalated cases at a reasonable review cost, the company can route only that bounded class to Sol. The remaining cases stay human-owned.

Use a prompt that exposes evidence and limits

A comparison prompt should make unsupported confidence visible. Adapt this compact version:

Task:
Allowed sources:
Required output:
Acceptance criteria:
Prohibited actions:

Use only the supplied evidence. Cite the exact source section for each material claim. If the evidence is missing or conflicting, say what is missing and stop. Do not send, publish, purchase, delete or change a live system. End with a reviewer checklist and a confidence note tied to the evidence.

Keep the core prompt stable across the comparison. If one model receives extra context, a different tool set or more reasoning effort, record the difference. Otherwise the result is not a fair routing decision.

How to treat OpenAI's benchmark claims

OpenAI published comparisons across business automation, coding, computer use and factuality. The announcement reports substantial gains and lower cost per task in several evaluations. Those results can help a team decide what to investigate, but they remain vendor-selected evidence produced with particular prompts, tools, effort settings and scoring methods.

Ask whether the benchmark resembles your real task, whether the configuration matches your account and whether a reviewer can inspect the output. A percentage improvement does not reveal the cost of a rare but serious failure in your workflow. Save accepted and rejected examples from your own pilot so later model changes can be compared against the same evidence.

Migration checks before changing a live workflow

• Confirm the exact model identifier, surface, account and date.
• Recheck data retention, training, regional processing and access controls for the product you use.
• Measure cache hit rate rather than assuming the maximum discount.
• Set token, retry and tool-call limits.
• Keep a fallback and a stop condition; do not silently switch models after a failed request.
• Log the prompt, evidence version, model and reviewer decision for consequential tasks.

A model release should trigger an evaluation, not an automatic migration. OpenAI's pricing and availability make Sol and Luna meaningful candidates, especially where repeated work makes cost visible. The durable decision is a routing policy backed by task evidence.

Reader Q&A

When were GPT-6 Sol and Luna announced?

OpenAI announced GPT-6 Sol and GPT-6 Luna on 22 September 2026. The official announcement and availability details in this article were checked on 23 September 2026.

How much do GPT-6 Sol and Luna cost in the API?

At announcement, OpenAI listed GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna was listed at $0.10 input and $0.50 output per million tokens. Confirm current pricing and platform fees before budgeting.

Are GPT-6 Sol and Luna available in ChatGPT?

OpenAI said they were available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna was also available to Free and Go users in the desktop app. The models were not yet available in Chat at announcement, and rollout was gradual.

Which model should a small business choose?

Start with the least expensive model that can meet the workflow's acceptance criteria. Test Luna for bounded high-volume tasks, Sol for harder multi-step work and Astra only where its added capability materially improves a consequential result.

Does lower token pricing guarantee lower workflow cost?

No. Total cost can include reasoning, uncached context, output, retries, tool calls, failed attempts and human review. Measure the cost of an accepted result rather than comparing token prices alone.

Should an existing workflow switch immediately?

No. Run a controlled comparison using representative non-sensitive tasks, the same evidence and explicit acceptance criteria. Approve a narrow route only after reviewing errors, total cost, permissions and stop conditions.

Make the model choice reviewable

Write down the task class, model identifier, evidence version, test date, observed cost and reviewer decision. That record turns a fast-moving product announcement into a change you can inspect, reverse and improve.

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio