Codex+Ona Cloud Agents vs Claude Code Background Tasks (June 2026)

OpenAI announced on June 11, 2026 it will acquire Ona (formerly Gitpod) to give Codex persistent cloud execution environments — agents that run for hours or days inside customer VPCs. Anthropic’s Claude Code already supports background sub-agents that run autonomously for shorter durations on Anthropic’s infrastructure. This page compares the two long-running coding agent approaches.

Last verified: June 14, 2026

TL;DR

• OpenAI Codex+Ona: Customer-VPC execution, hours-to-days, kernel-level security, regulatory approval pending. GA likely Q3–Q4 2026.
• Claude Code background sub-agents: Anthropic-cloud execution, minutes-to-hours, strongest reasoning on Opus 4.8 / Fable 5. Available now.
• Net: For regulated industries with strict in-house cloud requirements + long-running migrations, Codex+Ona is positioned to win. For deepest reasoning on hard agentic refactors today, Claude Code leads.
• Don’t oversell the “agents run for days” framing. The practical win is agents that survive overnight without supervision; multi-day autonomy is more aspirational than routine in mid-2026.

What Ona actually is

Ona is a cloud platform for running teams of AI software-engineering agents. Built on what was previously Gitpod (founded Kiel, Germany 2020). Core workflow: a developer assigns a task; Ona returns a pull request. Each agent gets a full cloud environment — not just a sandbox — with the customer’s tools, network access, and scoped permissions. Runs inside the customer’s VPC with audit trails and kernel-level policy enforcement.

Ona’s named customers (per ona.com platform page): BNY, Pearson, GSR, Vanta, EquipmentShare, Hargreaves Lansdown. One unnamed top-100 global company reports 90–95% of migration work completed by Ona Automations, 4x productivity increase, 83% of pull requests co-authored by Ona, 400+ Python repositories modernized in six months.

OpenAI’s stated rationale for the acquisition: agents need “more than intelligence — they need a trusted workplace” (Ona CEO Johannes Landgraf). Acquisition subject to regulatory approval.

For the full Ona explainer, see OpenAI Acquires Ona (ex-Gitpod): Codex Gets a Cloud and OpenAI Codex Cloud vs Anthropic Claude Cowork.

Side-by-side

Dimension	Codex + Ona (post-integration)	Claude Code background sub-agents (today)
Execution location	Customer VPC	Anthropic cloud
Trust boundary control	Customer	Anthropic
Agent duration	Hours to days	Minutes to hours
Audit trail	Kernel-level, customer-controlled	Anthropic-managed logs
Default model	GPT-5.5 (then newer Codex)	Opus 4.8 / Fable 5 (US-only)
Cost model	Compute on customer cloud + OpenAI model API	Anthropic credits (Pro $20/mo + metered)
Availability	Q3–Q4 2026 (regulatory pending)	Available now
Existing customer signal	BNY, Pearson, Vanta, GSR, etc. (Ona pre-deal)	Anthropic enterprise + dev shops
Best for	Long-running migrations in regulated industries	Deep agentic refactors today
Pull request output	Native	Native (Claude Code generates branch + PR)

When each one wins

Codex+Ona wins when

• Regulatory / compliance requires customer-controlled cloud execution. Banks, hospitals, defense contractors, certain government agencies — the trust boundary control is the killer feature.
• Tasks are long-running — multi-day migrations, large codebase modernizations, refactors that genuinely take 24–72 hours of execution time.
• You already standardize on OpenAI for model access, with ChatGPT Enterprise / Codex licensing.
• Audit trails must satisfy regulators with kernel-level visibility (financial services compliance, healthcare audits).
• You have a high-trust on-call rotation that can verify long-running agent output without close supervision.

Claude Code background sub-agents win when

• You need the deepest reasoning on hard agentic tasks today (Opus 4.8 / Fable 5 lead SWE-Bench Pro at 69.2%+).
• You want it now. Codex+Ona is not GA until Q3–Q4 2026; Claude Code sub-agents are shipping.
• Tasks are minutes-to-hours, not days. Most agentic work falls here.
• You don’t have hard customer-cloud requirements. Anthropic-hosted execution is fine for most teams.
• You’re already on the Anthropic stack for safety/compliance — Claude Code is the natural agentic surface.
• You want MCP-native integration with your tools (Anthropic authored the MCP spec; ecosystem is most mature).

The realistic limits of “agents run for days”

Both vendor narratives lean into the “agents that run for days” framing. Three real-world limits:

1. Cost compounds. Long-running agents accumulate context that compounds API costs. A multi-day task on frontier models can run $5–$50 in pure API spend even with Ona’s customer-cloud execution (Ona reduces compute cost, not model API cost). Plan budgets accordingly.

2. Drift is real. Agents running for days hit context-window limits and need checkpointing/summarization strategies. Production patterns for this (LangGraph, custom orchestration, vendor-managed checkpointing) are still maturing in June 2026. Expect 10–25% of long-running agent runs to fail or drift in ways that require human intervention.

3. Review velocity is the real bottleneck. Agents that produce pull requests need humans to review. If your engineering org’s PR review velocity is the constraint, agent execution time is irrelevant — you’re rate-limited downstream. Investing in PR review tooling and reviewer-friendly agent output is often higher-leverage than longer agent execution.

The practical win for most teams in 2026 isn’t “agents run for days.” It’s “agents survive overnight without supervision and have a coherent PR waiting in the morning.” That bar is achievable today with both Codex+Ona’s roadmap and Claude Code’s sub-agents.

Cost models

Codex+Ona (projected post-integration):

• Compute: runs on customer cloud (AWS / GCP / Azure). Customer pays cloud bill.
• Model API: OpenAI Codex pricing — currently per-token Codex model pricing.
• Ona/OpenAI orchestration fee: TBD post-integration.
• Expected enterprise pricing: per-seat + usage above included quota.

Claude Code background sub-agents (today):

• $20/mo Pro plan with included credits + metering above.
• $200/mo Max plan with much larger included pool.
• Per-task cost for a typical multi-hour background sub-agent run on Opus 4.8: $1–$10 depending on context and complexity.

For most teams, Claude Code is cheaper to start ($20/mo + metering vs Codex+Ona’s enterprise contract + cloud bill). Codex+Ona pulls ahead at large scale or when customer-cloud requirements force the choice.

The integration timeline

OpenAI’s Ona acquisition is subject to regulatory approval (US antitrust, EU merger control). Realistic timeline:

• Q3 2026: Regulatory review concludes (assuming no second request).
• Q3–Q4 2026: Closed-beta Codex+Ona integration available to enterprise customers.
• Q4 2026 – Q1 2027: GA rollout broader.
• 2027+: Native Codex CLI + Ona cloud execution becomes default for enterprise Codex deployments.

If your timeline requires production long-running agents in 2026, Claude Code sub-agents (or Cursor’s cloud Composer, or Devin Desktop) are the only available options. If you can wait 6–9 months, Codex+Ona is worth evaluating for regulated-industry use.

The decision framework

Question 1: Do you need agents running inside your own VPC for compliance?
  Yes → Codex+Ona (when GA) is the strongest fit
  No  → Continue to Q2.

Question 2: Do you need long-running agents in production today?
  Yes → Claude Code background sub-agents (or Cursor cloud, or Devin Desktop)
  No  → Continue to Q3.

Question 3: Are you doing hardest agentic refactors / multi-file work?
  Yes → Claude Code + Opus 4.8 (SWE-Bench Pro leader at 69.2%)
  No  → Continue to Q4.

Question 4: Are you OpenAI-standardized + can wait until 2027?
  Yes → Plan for Codex+Ona; pilot today with Codex CLI
  No  → Claude Code is the safer default

What to watch next 90 days

• Regulatory approval movement on the Ona acquisition — first procedural milestones (US antitrust filings).
• Anthropic’s response — likely Claude Code background sub-agent improvements + announcement of customer-VPC execution option (Anthropic has signaled interest).
• Cursor cloud Composer roadmap — Cursor 4.x has been rolling out cloud execution options; could become a third strong option.
• Devin Desktop long-running agent story — Cognition has been investing here; expect releases.

Related reading

• OpenAI Acquires Ona (ex-Gitpod): Codex Gets a Cloud
• OpenAI Codex Cloud vs Anthropic Claude Cowork (June 2026)
• OpenAI Ona vs E2B vs Modal: AI Agent Sandboxes
• Cursor 4 vs Codex CLI vs Claude Code: Which to Pick

---

Codex+Ona integration is subject to regulatory approval; timeline estimates here reflect current public guidance and may slip. Verify before architecting production dependencies.