OpenAI swaps GPT-4o for o3 inside Operator, leaves API unchanged

OpenAI is replacing the GPT-4o-based model that powers Operator with a version based on OpenAI o3, according to an addendum to the o3 and o4-mini system card published by the company. The change applies to the Operator product itself. The API version of Operator, OpenAI states plainly, "will remain based on 4o."

The disclosure is short, but it carries weight. Operator is OpenAI's agent that can take actions on a user's behalf in a web browser, clicking through pages, filling forms, and completing multi-step tasks. The model that sits underneath an agent like this determines how well it plans, recovers from errors, and resists manipulation on untrusted web pages. Moving the product from GPT-4o — a model released in mid-2024 — to o3, one of OpenAI's reasoning models, signals that OpenAI now considers its reasoning-line models mature enough to drive a consumer-facing agent in production.

The split between product and API is the detail developers will notice first. OpenAI has drawn a line: users of the Operator product get o3, while customers who access Operator through the API continue to get the GPT-4o-based version. That creates two different capability surfaces for what is nominally the same tool, and anyone building on the Operator API cannot assume their integration behaves like the consumer product.

The announcement arrives as an addendum to the system card for o3 and o4-mini, the safety and capability documentation OpenAI publishes alongside major model releases. System-card addendum}s typically record changes made after the original evaluation work, which fits here: the o3 and o4-mini card documented those models, and this note extends the record to cover Operator's new deployment. For researchers and policy watchers, that placement matters. It ties the Operator swap to OpenAI's existing safety documentation for o3 rather than standing it up as a separate, lighter-weight announcement.

The move also fits the broader trajectory of OpenAI's agent strategy. Since introducing its reasoning models, OpenAI has positioned them as the engines for agentic work, where a system must chain together decisions rather than answer a single prompt. Operator was originally launched on GPT-4o. Upgrading the product to o3, while holding the API back, suggests a staged rollout: OpenAI gets o3 into live consumer traffic first, with the API transition — if one comes — left for later. The company's statement does not say whether the API will eventually follow.

What the statement also does not include is equally telling. OpenAI has not published benchmark comparisons for Operator-on-o3 versus Operator-on-4o, has not named a date beyond the addendum itself, and has not detailed any new safety evaluations specific to the swap. The brevity leaves open questions that matter for anyone studying agent reliability: whether o3's reasoning changes Operator's failure modes on adversarial web pages, and whether the earlier safety testing documented for the GPT-4o-based Operator still describes the shipping product.

For the competitive picture, the change lands amid a rush of agent upgrades across the industry, with every major lab pushing stronger models into tools that act rather than chat. OpenAI's decision to document the swap through its system-card process — rather than a marketing post — is consistent with how the company has handled model changes it considers safety-relevant.

Developers should treat the API note as the operational takeaway: nothing changes for Operator API integrations today. End users of the Operator product are now interacting with an o3-based agent. Whether OpenAI converges the two surfaces, or keeps the API a generation behind its consumer agent, will be the signal to watch in future model and system-card updates.