GPT-6: Astra, Sol and Luna
Choose a GPT-6 model, check access and costs, then try a prompt or inspect a worked example.
Astra launched September 3; Sol and Luna joined on September 22. Compare the three models or read OpenAI’s latest family announcement.
What GPT 6 is
Latest Sol update: GPT-6.1 arrived September 29, 2026. Explore GPT-6.1 Sol pricing, API and access for the newer model’s settings.
GPT-6 is OpenAI’s model family: Astra launched on September 3, 2026; Sol and Luna followed on September 22. Astra serves demanding work, Sol focuses on coding and agent workflows, and Luna targets focused tasks at high volume. The model directory explains each option and links to its official reference.
Your choice depends on the task and the product. GPT-6 Pro in Chat uses Astra. Sol and Luna are available in Work and Codex, while Free and Go users can access Luna in the desktop app. API projects use the exact model IDs and separate billing. Use the access self-check to match a model to your plan, client, and country or region.
Find your GPT 6 starting point
| Your task | Start here | What you get |
|---|---|---|
| Pick a model | Astra, Sol and Luna | Model details, limits and relevant resources |
| Compare your workload | Model comparison | Shared capability checks and costs for the same token budget |
| Check availability | Access guide | Chat, Work, Codex and API instructions |
| Estimate a bill | OpenAI API pricing calculator | Request and monthly text-token estimates |
| Reuse long instructions | Prompt caching calculator | Read/write costs and the reuse needed to save money |
| Fit a prompt | Token counter | Text count, output reserve and remaining context |
| Start a task | Prompt library | 21 templates with inputs, outputs and checks |
| Inspect the work | Examples | Coding, documents, a playable game and structured data |
Which GPT-6 model should I try?
Start from a small task with a result you can check. For an expensive decision or a difficult project spanning several tools, review Astra. For a repository change with a failing test, try Sol. For repetitive classification or extraction, begin with Luna and validate the output schema.
The comparison tool keeps the same input and output counts across models. It shows published capabilities and token costs; choose your model after checking a representative result. A lower token price can save money on repeated calls, while retries or a larger reasoning budget can change the total.
Resources you can actually use
Each prompt has complete instructions, an output format and review checks. Copy prompt copies the reusable version. The worked example keeps its input separate from the answer, so you can rerun the task and compare the result.
For coding, follow the Sol pagination review: the fixture has deliberate defects, and executable assertions check the repair. For a repeated text task, open Luna support triage and compare the eight source requests with the structured response. The skills library packages repeatable procedures as downloadable SKILL.md files.
The existing Astra collection includes a frontend QA repair, document formatting, document reconciliation and a playable game. Original inputs, files and checks remain attached to those cases.
How our GPT 6 examples were run
The original 15 prompt examples and four workflow exercises used GPT-6 Astra at medium effort in Codex. The newer Sol and Luna cases record their own model, input, output, run date and checks. They use synthetic materials; the records describe those runs, rather than paid API or ChatGPT-interface tests.
To repeat an example, retain the source IDs and acceptance checks. For a code change, run the fixture before and after the repair. For extracted records, validate the schema and compare every field with the supplied text. This makes a useful answer easier to distinguish from a plausible one.
Follow GPT-6 updates
The release timeline separates the first family launch from later models. News records dated announcements; the pricing guide and API guide explain their current practical effect. Official facts carry source links and a verification date. Experiment dates stay with the original work.