OpenAI folds Codex harness into free Agents API, pressuring Anthropic

OpenAI has stopped treating its agent-orchestration layer as an internal advantage and turned it into a product, giving every API developer the same harness that already powers Codex and ChatGPT for Work. The move squeezes the margin available to startups selling agent infrastructure and puts pressure on Anthropic, whose own agent tooling now faces a free, fully managed rival from the model maker with the largest developer base.

The Agents API went into public beta at 8am Pacific on 10 September, alongside GPT-Live-1 and ChatGPT for Financial Services. OpenAI describes it as a way to "Build and run cloud agents with the Codex harness, fully managed by OpenAI." One API call now passes the task, model, tools and environment, and the harness handles sessions, orchestration, context compaction, tool search and subagents itself.

The architecture has four parts: an Agent that bundles the model, instructions, tools and MCP servers; an optional Environment sandbox; a durable Session; and a stream of events. Agents can run for hours or days. Sandboxes can sit on OpenAI's own infrastructure, in a customer's cloud, or with partners including Cloudflare, DigitalOcean, Oracle and Vercel — but OpenAI always runs the harness itself, keeping control of the layer that matters most commercially even when the compute sits elsewhere.

Pricing is the sharpest competitive signal: "There are no additional fees for using the Agents API," OpenAI said. Customers pay only for model tokens, tools and OpenAI-hosted container time. That removes a line item that agent-infrastructure vendors have been charging for and makes the harness a loss-leader for token volume rather than a standalone revenue product.

The beta carries real constraints: data storage is US-only for now, and Zero Data Retention isn't supported yet, which will hold back some regulated customers. Early adopters named by OpenAI span insurance, real estate, compliance and AI consulting — Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai and Nash.ai — suggesting the initial pull is coming from vertical software builders rather than pure AI labs. Details are in MarkTechPost's report.

Bundling the harness for free is a land-grab, not a pricing decision. OpenAI is betting that owning the orchestration layer — sessions, compaction, subagents — locks developers into its models even where third-party tools could substitute, and that token consumption from long-running agents more than offsets the lost harness revenue. Anthropic's agent tooling now has to compete against a zero-margin bundle from a rival with broader distribution, which raises the bar for differentiation on reliability or vertical fit rather than price. Startups selling standalone agent-orchestration layers are the most exposed: OpenAI has just commoditised the thing they charge for. The US-only data storage and missing Zero Data Retention support are the gaps to watch — whoever closes them first likely wins the regulated-industry customers still on the sidelines.