OpenAI has published the rate card for Codex, its autonomous coding agent, framing the economics of machine-driven software development. The model operates on a consumption-based system using 'compute units,' a significant departure from the per-seat subscriptions common with tools like GitHub Copilot.

Each Codex task spins up a cloud sandbox where the agent reads repos, writes code, executes it, and iterates. This closed-loop execution is resource-intensive: a task using the o3 model costs 50 units, while the lighter o4-mini costs 25. ChatGPT Pro subscribers receive 1,000 units monthly, allowing only 20 o3-level coding tasks before exhausting their budget.

For engineering teams, this architecture forces immediate trade-offs. The cost scales directly with activity, aligning expense with usage but introducing budget unpredictability. It also creates an internal resource allocation problem, as these units are shared with other advanced ChatGPT features, forcing a choice between automated coding and other reasoning tasks.

The pricing signals that OpenAI treats agentic coding as a premium, computationally expensive service. The sandboxed execution environment, which tests and verifies code, demands more than simple inference, justifying the higher unit cost. However, this model places the burden on teams to constantly evaluate if the compute cost for a task, especially one requiring multiple iterations, justifies the saved developer hours. Failed or subpar outputs that require significant human correction directly undermine the value proposition.

This move sets a benchmark in a competitive field. Whether consumption-based pricing becomes standard hinges on a simple, practical metric: the cost per *usable* outcome, not just per task.

Source: Webpronews