OpenAI Acquires Ona: Enterprise Persistent Agents and What They Cost

What OpenAI Just Acquired and Why

On June 11, 2026, OpenAI announced the acquisition of Ona, a company that provides secure persistent cloud environments. The motivation is clear: over 5 million developers now use Codex weekly — a 400% increase — and the current architecture cannot keep up with the demand for long-running agent tasks.

Ona gives Codex something it did not have before: persistent environments where agents continue working for hours or days, even when the developer's laptop is closed. This enables enterprise deployment with customer-controlled cloud infrastructure rather than OpenAI-managed compute.

The acquisition signals a fundamental shift in how AI coding agents are deployed — from ephemeral sessions to persistent, always-on compute environments. This shift has major cost implications.

Session-Based vs Persistent Agent Costs

Today's AI coding tools — Cursor, Claude Code, GitHub Copilot — operate on a session-based model. You interact, the agent works, the session ends. You pay for tokens consumed during active use. When you close your laptop, spending stops.

Persistent agents flip this model. The agent keeps working after you disconnect. This means:

• Compute runs continuously: Cloud environment stays provisioned whether or not you are watching
• Token generation continues: The agent thinks, writes code, runs tests, iterates — all billable
• Infrastructure costs accrue 24/7: Persistent VMs/containers cost money even during idle periods

What Persistent Agents Actually Cost

Let us estimate what a persistent Codex agent costs for a typical enterprise task — say, implementing a feature that takes 4 hours of agent work overnight:

• Token costs: 4 hours of active reasoning at ~50K tokens/hour output = 200K output tokens. At GPT-equivalent pricing (~$10/M output): $2.00
• Infrastructure: Persistent cloud environment with IDE, terminal, file system. Estimated $0.50-2.00/hour = $2.00-8.00 for 4 hours
• Total: $4.00-10.00 per overnight feature implementation

Compare this to the session-based alternative: a developer interacting with Claude Sonnet 4.6 for the same task would spend 2-3 hours of active collaboration, consuming perhaps 100K output tokens at $15/M = $1.50 in tokens with no infrastructure cost. But they also spend 2-3 hours of their own time.

The persistent agent costs 3-7x more in AI spend but frees the developer completely. For a senior engineer costing $100+/hour, the $10 agent cost is negligible.

Enterprise Budget Implications

For enterprises adopting persistent agents at scale, the cost model changes fundamentally:

The 3-5x cost increase is justified only if persistent agents deliver proportional productivity gains. Early Codex data suggests they do — tasks that previously required synchronous developer attention can now run overnight, effectively multiplying engineering capacity.

How This Compares to Anthropic and Cursor

Anthropic's approach with Claude Code remains session-based — powerful agents that work while you are present but stop when you disconnect. This keeps costs predictable and lower but limits throughput to developer-attended hours.

Cursor occupies a middle ground with background agents that can continue short tasks, but not multi-hour persistent work. Their Auto-Review feature (launched the same week as the Ona acquisition) focuses on making session-based agents more autonomous rather than making them persistent.

The competitive landscape is splitting into two camps:

• Persistent (OpenAI/Codex): Higher cost, higher throughput, agent works while you sleep
• Session-based (Anthropic/Cursor): Lower cost, developer-attended, more predictable spending

What Developers Should Budget For

If your team plans to adopt persistent agents through Codex, budget for the infrastructure layer in addition to token costs. The Ona acquisition means customer-controlled cloud infrastructure — you will likely provision your own compute through AWS, GCP, or Azure, with Codex orchestrating on top.

For most teams today, the session-based model with Claude Sonnet 4.6 at $3/$15 per million tokens or DeepSeek V4 Pro at $0.435/$0.87 remains the more cost-effective choice. Persistent agents make economic sense only when the value of overnight autonomous work clearly exceeds the 3-5x cost premium — typically for teams where developer time is the bottleneck, not AI spend.