OpenAI Japan maps Codex work across GPT-6 Astra, Sol and Luna

The starting problem is cost control: running a top model through every step of development strains budgets, while defaulting to a cheap model multiplies correction loops and rework. Ishida's answer is not a single model choice but a routing discipline — reserve GPT-6 Astra for demanding tasks, use GPT-6.1 Sol where quality and speed must balance, and hand routine work to GPT-6 Luna.

The examples cited show what that discipline can look like in practice. Rakuten shortened a mobile app build that had taken about three months to a few weeks, and an OpenAI internal project produced roughly 1,500 merged pull requests over about five months, or 3.5 per person per day. In both cases, individual development tasks were delegated to AI while people focused on higher-level decisions, which is the productivity gain Ishida points to.

What the guidance stresses is that Codex does not deliver results on its own. Teams also need to prepare environments and rules suited to their work, and decide which model handles which task — the difference between simply handing off processing and actually improving development output.