OpenAI is now running the agent plumbing on developers’ behalf, and not charging for the plumbing itself. Its Agents API, in public beta since 10 September, is a managed service built on the same Codex harness that sits behind Codex and ChatGPT, so developers hand over session orchestration, context compaction and recovery rather than building those themselves — the part of agent engineering that tends to consume the most effort.
The API supports tool search and programmatic tool calling, multi-agent orchestration with parallel subagents, MCP and custom functions, and either OpenAI-hosted sandboxes or self-hosted compute through partners including Modal, Cloudflare and E2B. There are no beta fees; usage is billed as tokens and tools consumed.
The same day, OpenAI made GPT-Live-1 generally available, priced at $0.05 per minute of voice session with the underlying model and tool usage billed separately, and began letting administrators enforce maximum API key lifetimes at organisation or project level.
OpenAI’s post quotes customer figures — an evaluation score moving from 0.71 to 0.85, a 60 per cent cut in cost per case — that are supplied by those customers and not independently verified.
Why it matters: the managed harness is where the agent platform contest is now being fought, and charging nothing for the orchestration itself sets a reference point competitors will have to answer.