OpenAI Launches Workspace Agents and GPT-Rosalind Model

On April 23, 2026, OpenAI announced the release of Workspace Agents in research preview, alongside the specialized GPT-Rosalind model for life sciences. The twin releases represent a structural transition for the platform. Standard conversational assistance is being supplemented by persistent, domain-specific autonomous partners.

Persistent Automation with Workspace Agents

Workspace Agents execute multi-step workflows across shared team environments. Unlike standard ChatGPT sessions, these agents operate continuously in the cloud. They remain active and process tasks when users are offline.

The architecture relies on the Codex model for execution. The agents write and run code, interact with third-party applications, and maintain a shared memory state that updates based on team feedback. At launch, the system supports over 60 applications, including Slack, Google Drive, Microsoft Office suites, Salesforce, and Jira.

OpenAI positions Workspace Agents as the direct successor to Custom GPTs. A conversion tool will be provided to migrate existing setups into the new agent architecture. This shifts the paradigm from basic custom prompting to continuous background execution. If you build internal tooling, this requires a different approach to understanding how AI agents operate and executing tasks over long horizons.

Access is currently limited to ChatGPT Business, Enterprise, Education, and Teacher plans. Usage remains free during the preview period until May 6, 2026. Following the preview, OpenAI will implement a credit-based pricing model similar to standard API billing.

GPT-Rosalind Benchmarks and Capabilities

GPT-Rosalind is a frontier reasoning model optimized specifically for biochemistry, drug discovery, and translational medicine. It is designed to handle long-horizon, tool-heavy scientific workflows like protein engineering and genomic analysis.

The model demonstrates measurable gains over general-purpose counterparts in domain-specific evaluations. On the BixBench benchmark, GPT-Rosalind achieved a Pass@1 score of 0.751. This outperforms GPT-5.4 at 0.732, Grok 4.2 at 0.728, and Gemini 3.1 Pro at 0.550. On LABBench2, GPT-Rosalind outperformed GPT-5.4 on six out of eleven tasks. It reached the 95th percentile in RNA sequence prediction.

The model ships with a free Life Sciences research plugin for Codex. This integration connects researchers to over 50 scientific tools and databases, spanning human genetics, protein structure, and clinical evidence. GPT-Rosalind is currently in a trusted access research preview for qualified enterprise customers. Launch partners include Amgen, Moderna, Thermo Fisher Scientific, and the Allen Institute.

Governance and Enterprise Controls

Continuous execution introduces new requirements for access control. Administrators can set approval checkpoints for sensitive actions. This ensures a human in the loop before an agent sends external emails or edits financial spreadsheets. Tool access permissions can be scoped globally or per team across the organization.

The platform competes directly with Microsoft Agent 365, Salesforce Agentforce, and Anthropic Claude for Work. If you deploy these systems, you will need strategies to manage API costs as continuous background processing consumes credits far faster than manual chat interactions.

Development teams should inventory their existing Custom GPTs to prepare for the migration tool. Transitioning from synchronous chat to asynchronous agents requires updating your internal governance policies and defining clear boundaries for automated system interactions.