OpenAI upgrades Operator with o3 for smarter, safer tasks

OpenAI is rolling out a significant upgrade to Operator, its autonomous AI agent capable of browsing the web and interacting with cloud-hosted software environments. The company has announced that Operator will soon run on a new version of its o3 model—part of the latest o-series, known for superior reasoning and mathematical capabilities.

Previously powered by a custom version of GPT-4o, Operator's transition to o3 represents a strategic leap in performance and reliability. According to OpenAI, o3 outperforms GPT-4o on numerous benchmarks involving logic, math, and decision-making. The API-based version of Operator, however, will continue to use GPT-4o for now.

This move is part of a broader trend in the AI industry: the race to build agentic AI—tools that can carry out complex digital tasks with minimal supervision. Companies like Google and Anthropic are pushing similar initiatives. Google’s Gemini API supports a “computer use” agent and the more consumer-facing Mariner, while Anthropic’s models are capable of performing system-level tasks like file handling and web navigation.

The new version, branded o3 Operator, has been fine-tuned with additional safety protocols tailored for computer use. These include curated datasets that help the model learn clear boundaries around confirmations, refusals, and responsible behavior. This tuning is essential for agents with real-world capabilities, where trust and reliability are paramount.

OpenAI has also released a technical report detailing how o3 Operator compares to its predecessor on key safety metrics. Notably, o3 Operator is less prone to carrying out restricted actions or leaking personal data, and it shows greater resistance to prompt injection attacks—a method where malicious users trick AI models into bypassing safety protocols.

Despite inheriting the powerful coding abilities of o3, OpenAI confirms that o3 Operator does not have direct access to a coding environment or terminal, reinforcing guardrails for safe deployment. The model uses a multi-layered safety approach, continuing the protective standards established with GPT-4o.

As the AI arms race intensifies, the enhanced Operator positions OpenAI at the forefront of agentic AI—paving the way for more capable, safer digital assistants that can autonomously fulfill increasingly complex user demands.