OpenAI

OpenAI told the world on 11 June that it was buying Ona, a German company better known until recently as Gitpod. The deal folds Ona’s cloud-execution technology into Codex, OpenAI’s coding assistant, and the stated goal is blunt: let coding agents keep working for hours after a developer shuts the laptop and walks away. Neither side disclosed what OpenAI paid. The acquisition lands in the middle of a year that has already reshaped how AI companies compete on code. SpaceX bought Cursor for $60 billion . Cognition’s valuation hit $26 billion on the back of Devin, its autonomous coding agent .

Now OpenAI is betting that the next fight isn’t over which model writes the best function, but over which company controls the infrastructure an agent runs on once nobody is watching it work. For engineering teams in Ireland, where US tech multinationals employ a large share of the software workforce, the deal matters less for its price tag than for what it signals. Cloud execution for AI agents is turning into its own product category, one that could reshape how companies buy and govern coding tools over the next year. Don't miss new tech stories on Google Add Tech Insider once in the Google app and our stories appear in your news suggestions.

” Bloomberg first reported the transaction, and Forbes and other outlets followed within two days. Every account of the story includes the same caveat: neither company shared financial terms, and the deal still needs regulatory approval before it closes. Once it does, OpenAI says, Ona’s team moves directly into the Codex group rather than operating as a separate subsidiary. Ona is the rebranded identity of Gitpod, a German cloud-development company built around a simple pitch: open a browser and a fully configured coding environment appears, no local setup required.

That idea made sense for human developers tired of “it works on my machine” bugs. It also happens to describe almost exactly what an autonomous coding agent needs. ” Tech Times characterized the acquired technology in similar terms: cloud-based sandboxes where AI agents can reach tools, systems, and project context without a live connection back to someone’s desktop. The rebrand from Gitpod to Ona happened before the OpenAI deal was announced, tracking a broader pivot in the company’s product story from human-facing cloud IDEs toward agent infrastructure.

The detail that keeps surfacing in coverage of the deal is persistence. A chat-based coding assistant answers one prompt at a time and waits. An agent tasked with migrating a codebase to a new framework needs to run for hours, sometimes across a weekend, testing changes and correcting course without a developer supervising every step. The Agentic Review’s analysis of the deal described Ona’s approach as running agents inside short-lived cloud sandboxes with automatic checkpointing every few minutes, so an interrupted session can resume instead of restarting from zero.

OpenAI has not confirmed that specific implementation detail, so it should be read as informed reporting rather than an official spec sheet. What OpenAI has confirmed is the goal: agents that survive a closed laptop, a dropped VPN connection, or a task queue that runs longer than a single workday. The table below pulls together the verified details of the transaction as reported across multiple outlets, alongside the figures that remain estimates rather than confirmed facts. Codex has had a mixed run in 2026. OpenAI’s coding assistant crossed 5 million weekly users, a six-fold jump , showing real pull with developers even as OpenAI had to scrap its own SWE-Bench Pro benchmark after roughly 30% of its test tasks turned out to be broken .

That combination, fast growth alongside a credibility stumble on benchmarking, helps explain why OpenAI is now buying infrastructure rather than only shipping new model checkpoints. Owning the execution layer lets OpenAI make claims that are harder to fake than a leaderboard score: an agent that keeps a multi-hour refactor running securely inside a company’s own cloud account is a different sales pitch than a chat window that writes a function on request. Digital Applied’s analysis of the deal frames it exactly that way: OpenAI is buying the substrate underneath the agent, not just another feature bolted onto Codex.

Forbes contributor Janakiram MSV, writing days after the announcement, framed the acquisition as OpenAI’s move to run Codex agents directly inside enterprise clouds rather than asking IT departments to trust an external service with long-running, unsupervised code changes. That distinction matters to any large organization with strict data-residency or audit requirements, categories that cover much of Ireland’s multinational tech and financial-services sector. The pitch is straightforward: instead of a coding agent phoning home to a third-party server for hours at a time, the execution happens inside infrastructure the customer already controls and can audit.

Whether that pitch converts into signed enterprise contracts is a separate question, but it’s a coherent response to a market where trust, not raw model quality, has become the bottleneck for adoption. Ona is one of three very different acquisitions and valuation events that have defined the coding-tool business in 2026, and comparing them says a lot about where the money is actually going. SpaceX’s $60 billion purchase of Cursor bought a finished product with a large existing user base, an editor developers already open every day. Cognition’s $26 billion valuation was built on Devin’s ability to complete coding tasks with minimal supervision, a bet on agent capability itself.

OpenAI’s purchase of Ona is different in kind from both. It isn’t buying a developer-facing product or a headline benchmark result. It’s buying the plumbing: the sandboxes, checkpointing, and orchestration that let any agent, regardless of which model powers it, keep running once the human steps away. That’s a much smaller, much less visible acquisition by disclosed price, but it targets a layer of the stack that every other coding-agent vendor also needs to solve. Neither OpenAI nor Ona published a purchase price, which is common for deals of this size when the buyer is privately held and the target is small relative to the acquirer.

Into that gap stepped IDC research vice president Arnal Dayaratna, who estimated Ona’s 2025 revenue at roughly $7 million and used that figure to suggest a deal value somewhere in the $450–500 million range, according to Digital Applied’s reporting on the transaction. That estimate is exactly that: an outside analyst’s back-of-envelope math based on a revenue guess, not a confirmed transaction value. It bears repeating because coverage of acquisitions like this one tends to flatten an analyst estimate into a headline number within a few news cycles.

Until OpenAI or Ona files something with a regulator that states a price, the honest answer to “how much did this cost” is that nobody outside the two companies knows. Cloud development environments aren’t new. Gitpod, GitHub Codespaces, and Replit all built businesses in the years after 2019 on a version of the same pitch: skip the local setup, run your dev environment in someone else’s data center, and pick up work from any machine. The sales pitch back then was convenience for human developers who wanted to stop debugging broken local toolchains.

What’s changed in 2026 is the customer. The same category of infrastructure that once existed to save a developer twenty minutes of environment setup is now being repositioned as the substrate that lets an autonomous agent work unsupervised for a full day. Gitpod’s rebrand to Ona, ahead of this acquisition, tracks that shift almost exactly: same core technology, a different buyer in mind. Ireland’s position as a European base for US tech multinationals means decisions made in San Francisco about enterprise AI tooling tend to land on Irish engineering teams faster than in most of the rest of Europe.

A deal explicitly framed around deploying agents “securely in production” inside enterprise cloud accounts is aimed squarely at the kind of large, regulated organizations that dominate Dublin’s tech and financial-services employment base. The regulatory approval clause attached to the deal is also worth watching locally. A transaction that folds a German company into a US AI firm sits inside the EU’s merger-review process, and European regulators have grown more attentive to AI infrastructure consolidation over the past two years. How quickly this particular deal clears that process will say something about how EU regulators plan to treat smaller, infrastructure-layer acquisitions, which are becoming common in AI, as distinct from the model-level mega-deals that draw most of the headlines.

Handing an agent hours of unsupervised runtime raises the exact governance questions that have already surfaced elsewhere in the AI coding space this year. GitLab’s research found that 92% of firms can’t properly govern the AI-generated code already in their pipelines , and a separate industry study found that AI-written code carries roughly 70% more bugs than human-written code, with developer trust in AI output down to 33% . Persistent execution doesn’t cause those problems, but it raises the stakes on them. An agent that runs for ten minutes and stops for review is easy to audit.

An agent that runs for six hours across a codebase, checkpointing and resuming along the way, needs a much more deliberate approach to logging, rollback, and human sign-off before most security teams will let it near production systems. Teams already using Codex don’t need to act on this news immediately, since the deal hasn’t closed and no product changes have shipped yet.