OpenAI Acquires Ona: What Persistent Cloud Execution Means for Your Agent Routing and Cost Architecture

OpenAI announced on June 11 that it will acquire Ona, the company behind secure cloud development environments that have served over 2 million developers. The goal: give Codex agents persistent, customer-controlled environments where long-running work continues even after a laptop closes. For teams managing API gateways, multi-provider routing, and agent cost governance, this acquisition signals a structural shift in how agent workloads are priced, scoped, and observed.

What happened

OpenAI will fold Ona's cloud execution and orchestration technology into the Codex ecosystem. Ona's core product enables developers to work in secure, reproducible cloud environments — essentially moving the entire development workspace off local machines and into the cloud.

Codex now serves over 5 million weekly users, up 400% from earlier this year, and its most valuable tasks increasingly run for hours or days rather than minutes. The acquisition addresses a fundamental gap: agents need a place to keep working when the user walks away.

After closing, the Ona team will join OpenAI and integrate with the Codex team. Ona's customer-controlled execution model means agents operate inside the customer's own cloud environment while OpenAI provides the intelligence and orchestration layer. This is a split-plane architecture — compute lives in the customer's infrastructure, model inference lives in OpenAI's.

Why it matters for AI engineering teams

The shift from per-request to persistent agent sessions is already underway across the industry. OpenAI moved container sessions to per-minute billing with a 5-minute minimum on June 2. Anthropic launched self-hosted sandboxes for Claude Managed Agents in May. The Ona acquisition makes the same bet at platform scale: agents will work for hours, not seconds, and the cost model must follow.

Three things change immediately for teams running API routing or agent orchestration:

Session lifecycle gets longer and more complex. When an agent runs for 6–24 hours inside a persistent cloud environment, your routing layer must track not just individual API calls but entire work sessions. Idle detection, timeout policies, and cost ceiling enforcement move from "nice to have" to critical infrastructure. A Codex agent that spins up a persistent environment and runs overnight without guardrails can burn through a budget silently.

Credential scoping becomes a governance surface. Ona's customer-controlled model means the agent has access to the customer's cloud environment — their repos, their CI/CD, their secrets. The routing and gateway layer must enforce: which credentials are scoped to which agent session, how long they're valid, and what happens when a session is terminated. This is the same pattern Claude Managed Agents addresses with vault env-var credentials, but the OpenAI implementation will carry different blast-radius characteristics.

Cost attribution splits across two planes. You now have model inference cost (OpenAI's charges for GPT-5.5, Codex) and execution environment cost (the customer's cloud bill for the persistent runtime). Traditional API billing dashboards that track cost-per-token will miss half the picture. Teams need a unified view that correlates token spend with compute-hours, or they'll be surprised by the infrastructure bill the same way they were surprised by the OpenAI container session per-minute change.

The router/operator angle

This acquisition reinforces a pattern every API gateway operator needs to internalize: the unit of work is shifting from "request" to "session."

When agent work spans hours, your routing policy can't just optimize for latency on the next token. It must manage:

• Provider affinity over time. A persistent agent session should route to the same provider to maintain context. If your gateway load-balances across providers for cost or latency, you need session-aware routing — or you'll fragment the agent's context and degrade quality.
• Idle cost control. A persistent environment that the agent isn't actively using is still burning compute. Your gateway should surface idle detection and enforce auto-suspend policies.
• Credential rotation and revocation. Long-running sessions mean long-lived credentials. The security model must support mid-session credential rotation without killing the agent's work.
• Observability across planes. Traditional API metrics (requests/second, p95 latency, tokens/request) don't capture what matters for persistent agents: total session cost, idle time ratio, tool invocations per session, and environment lifecycle events.

This is structurally similar to the Claude Managed Agents self-hosted sandbox pattern we covered earlier — but OpenAI's version ties execution directly to the Codex product surface, which means the blast radius of a misconfigured gateway is larger: the agent can act on code, infrastructure, and data, not just generate text.

What TheRouter users should watch or try

• Review your session timeout and cost ceiling policies. If you're routing Codex or agentic workloads, ensure your gateway enforces hard cost ceilings per session, not just per request. Persistent agents that run overnight without caps are a billing risk.
• Plan for split-plane observability. Model inference cost and environment execution cost will live in different billing systems. Consider how your routing layer can surface both in a single dashboard.
• Audit your credential scoping for agent sessions. Ona's customer-controlled model means the agent will have broader access than a stateless API call. Ensure your gateway enforces per-session credential isolation and supports rotation.
• Watch for Ona integration landing in the Codex API. The acquisition is subject to regulatory approval, but once closed, expect new API parameters for environment lifecycle management. Your routing layer should be ready to handle persistent session metadata.