OpenAI PBC today unveiled a new platform called Frontier, designed to enable businesses to build and manage AI agents.
The product launch coincides with the release of the company's latest programming-optimized model. The new algorithm, dubbed GPT-5.3-Codex, not only surpasses its predecessor in code generation but also demonstrates superior performance in general productivity tasks.
Frontier empowers employees to create AI agents within a chat interface similar to ChatGPT. Users start by providing a natural language description outlining the agent's intended task and methodology. Subsequently, these agents can be integrated with various applications to automate manual workflows. According to OpenAI, Frontier offers connectivity to customer relationship management platforms, data warehouses, and a wide range of other services.
Employees have the option to enhance their agents with custom skills. These user-created extensions improve the agents' ability to handle multi-step processes. For instance, a skill could include pre-configured Kubernetes scripts that an agent can deploy to spin up container clusters.
OpenAI states that agents connected to Frontier "develop memory" to log their executed tasks. This memory enables them to progressively improve the quality of their outputs over time. Administrators can monitor agent performance via a dashboard that visualizes key metrics, such as the volume of customer support tickets processed and their resolution success rates.
Frontier incorporates additional observability features. The platform generates an audit log for every task completed by an agent, facilitating quality assessment. For example, an audit log for a technical support interaction might include visualizations quantifying the agent's politeness and the accuracy of its responses.
OpenAI plans to offer customers dedicated deployment engineers to assist in developing Frontier best practices. Furthermore, the company will collaborate with other AI firms to help them build Frontier-based functionalities. The initial roster of partners includes sales automation provider Clay Labs Inc., medical notes specialist Ambience Healthcare Inc., and several other startups.
The coding-focused model GPT-5.3-Codex, launched alongside Frontier today, boasts a 25% faster response time than its predecessor. This algorithm has set new records on the SWE-Bench Pro and Terminal-Bench programming benchmarks. The former evaluates coding problems across four programming languages, while the latter assesses an AI model's proficiency in using command-line interfaces.
OpenAI indicates that GPT-5.3-Codex is also suitable for a range of other use cases. The company evaluated the model using the OSWorld benchmark, which encompasses a broad set of online research and document editing tasks. GPT-5.3-Codex achieved a score of 64.7%, representing a 26.5% improvement over the previous model.
The model is available within the paid version of ChatGPT and will soon be accessible via OpenAI's application programming interface. Meanwhile, Frontier is currently accessible to a limited group of enterprise clients, including Oracle Corporation, Hewlett Packard Enterprise, and several other major technology firms. OpenAI intends to make the platform more widely available in the coming months.
