Enterprise

OpenAI opens its Agents API to every developer, putting the Codex harness one call away

Public beta hands small-team B2B builders the same session state, context compaction, and subagent coordination that runs Codex — with no fee beyond token and container usage.

By Asher Glick· Enterprise software desk·2026-09-11·3 min read

Photo: Unsplash / Chris Ried — Developer's laptop screen filled with source code

OpenAI opened its Agents API in public beta on September 10, 2026, exposing the same session, context-compaction, and subagent machinery that has been running Codex behind a single OpenAI-Beta: agents=v1 header. There’s no additional fee; developers pay only for the tokens and tools their agents consume.

The unbundling is the story. Codex has quietly grown into a platform for agentic software work, with heavy users averaging over 60 hours of agent operation per day as of June 2026. What OpenAI has now done is peel the harness off that product and rent it to everyone else. MarkTechPost’s read of the docs organizes the surface around four concepts: Agent, Environment, Session, and Events and items. Sessions run in four steps: create with a task, follow progress via streaming or webhooks, continue with a new task, or steer the current turn.

The quickstart runs against gpt-6-astra. Environments come in three flavors: OpenAI-hosted, self-hosted via codex exec-server, or one of nine partner sandboxes (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel). Tool search loads definitions only when needed, preserving the model’s cache. Subagents are switched on with multi_agent: { enabled: true, max_concurrent_subagents: 3 }, each holding its own context.

Vendor-supplied numbers are, per both MarkTechPost and AI Mastery, not independent. Ciridae reports its evaluation score climbing from 0.71 to 0.85 with a 4x latency reduction on subagent flows. SafetyKit claims 60% lower cost per case after migrating its case-review workflow. Hypha reports 86% fewer failed agent responses after separating harness from sandbox. Nash.ai says it’s running thousands of long-running agents across global logistics networks.

Two caveats sit in the release notes: the API is US-only, and it isn’t eligible for zero data retention even on a self-hosted sandbox. OpenAI paired the launch with a new Data agent inside ChatGPT Work that, per Seeking Alpha’s Brandon Evans, connects to company data and builds interactive dashboards.

The structural move mirrors what Salesforce’s Agentforce ARR ramp and Clay’s $7B GTM valuation have already priced in: the orchestration layer is where the margin lives, and OpenAI has decided to be the landlord.

Sources