OpenAI's cheaper GPT-6 tiers arrive, availability splits by plan

The Codex change is small and worth reading closely: a rate-limit prompt that recommends Luna means a model swap can happen mid-session, at the moment your usage is already under pressure. Pin model ids explicitly in Codex configuration and in any CI path, so a switch is a decision rather than a fallback. Two other things decide whether the lower list price reaches your bill. Cache discounts only pay when a prompt prefix stays byte-stable, so find every place you rewrite a system prompt per request. And the published numbers came from vendor harnesses, not your repository — before moving a coding agent off its current model, replay a set of your own recorded tasks and compare outputs and token counts, not headline scores.

This is an eligibility and jurisdiction question before it is a cost question. The plan matrix differs by surface: GitHub lists different Copilot plans for Sol and for Luna, and GitHub's changelog says new models are enabled automatically unless an administrator has turned off the global default — so deciding who reviews new models before staff can select them is a decision you make now or have made for you. Region handling also differs per model and deployment type in Microsoft's listing, with priority processing for Sol confined to US data zones, which matters if an EU data-zone commitment is written into a customer contract. And a lower price per token is not a lower price per task; agent runs that retry more will consume more.