The problem: cost per token and quality of work
For companies considering the adoption of artificial intelligence agents, the sticking point is the cost of using APIs, tied to tokens consumed. This is emphasized by an article in ITmedia AI+, republished by Keyman's Net, devoted to Codex, OpenAI's coding agent. The dilemma described is simple: running the most powerful model across all phases of the project squeezes the budget, while choosing an inexpensive model everywhere multiplies correction cycles and rework. The source's summary is that development speed and quality should be sought within a defined budget, not outside it.
Three models, different tasks
Codex allows choosing among three models: GPT-6 Astra for the most complex tasks, GPT-6.1 Sol for the best balance between quality and speed, and GPT-6 Luna for repetitive, low-cost tasks. Shuuhei Ishida, an applied AI engineer at OpenAI Japan, sets out a precise allocation: Sol for design, review, and revisions based on review; a switch to Astra when Sol is not enough; Luna for long implementations and high volumes of standardized activity. The stated lever is the combination of the three models throughout the development cycle, together with adjusting the amount of reasoning requested from the model.
The source also cites concrete results: in DeepSWE, a software engineering benchmark, Astra would be strong on the most complex tasks, while Sol would achieve similar accuracy at about one-fifth of the cost. According to Ishida, greater token efficiency reduces cost per task and shortens response times. These are company statements, not independent verifications.
From chat to written rules
Beyond model choice, the source insists on the organization of the environment. Codex is accessible from the terminal, integrated development environments, and a desktop application, with a common foundation, Codex Core, that manages the agent loop, tool execution, configuration, and authentication. It can also be invoked automatically via Codex Exec, the Codex SDK, and the Agents API.
On the operational side, Ishida proposes three tools. The first is a prompt template that makes explicit the objective, context, constraints, and completion conditions: without the latter, he warns, the agent may close out the work without verifying that it works, whereas indicating “continue until the build” keeps it active until the build passes. The second is plugins, which connect Codex to chat, email, shared files, and management tools. The third is AGENTS.md files and Skills: the former is read automatically at every startup and contains the repository structure, build and test commands, rules, and constraints; the latter stores reusable procedures that can be shared within the team.
What remains to be verified
The source reports cases cited by OpenAI, such as reducing from about three months to a few weeks the time needed to bring a web service to mobile, and about 1,500 change requests created and merged in five months in an internal project. These are examples presented by the company, without contextual data or measurement methodology. The picture is therefore an operational proposal based on the vendor's experience, not an independent assessment of the savings that can be achieved or of the thresholds beyond which it is worth switching from one model to another.
