GPT-6 expands beyond Astra
On September 22, OpenAI introduced GPT-6 Sol and GPT-6 Luna as faster, lower-cost additions to the family. Its launch positions Astra for the most demanding work and the new models for a broader range of everyday tasks.
The announcement covers the API, ChatGPT Work, and Codex. It names Plus, Pro, Business, Enterprise, and Edu access in Work and Codex, with desktop-app access to Luna for Free and Go users. It explicitly says Sol and Luna were not yet available in standard Chat at launch. That product distinction helps explain why different users may see different model menus.
The published API rates
The September 22 API changelog gives the following standard rates per million tokens for prompts containing up to 272K input tokens:
- GPT-6 Sol: $2 input, $0.20 cached input, and $10 output.
- GPT-6 Luna: $0.10 input, $0.01 cached input, and $0.50 output.
Longer prompts, cache writes, and other processing tiers have their own pricing details. A subscription's included usage is a separate question from these API rates. When budgeting, use the applicable rate category and measured usage rather than treating the headline output price as the cost of the whole task.
Prompt caching is part of the update
In a separate September 22 caching announcement, OpenAI described better reuse of shared context, a monitoring dashboard, and diagnostics for cache misses. It also introduced controls for choosing reusable prefixes and adjusting reasoning or tool availability while preserving reusable context.
For a website assistant that repeatedly receives the same publishing rules, our interpretation is that caching deserves attention alongside model selection. Stable shared material and frequently changing task content have different roles. A useful operational check is whether a prompt or tool configuration change unexpectedly raises the amount of freshly processed input.
What to try before changing a default model
Start with work whose expected result is clear: extracting fields from a supplied document, editing an existing article, or fixing a reproducible bug. Keep the material and acceptance rules consistent across candidates, and record both completion quality and actual consumption. Reserve more complex cases for a second comparison once the ordinary workflow behaves reliably.
This is our suggested rollout method, not an independent performance ranking. The relevant choice depends on your error tolerance, review effort, and task mix. Our earlier GPT-6 efficiency guide explains how to measure those tradeoffs.
Readers evaluating image-based work should also note the subsequent September 25 image-understanding fix before relying on their earliest test results.