On July 9, 2026, OpenAI launched ChatGPT Work, an autonomous agent that does not hand the work back to you. It accepts a goal, connects to your apps and files, breaks the job into steps, runs independently for hours, and returns a finished artifact: a completed spreadsheet, a slide deck, a document, or an interactive web app.
This is a meaningful shift in what enterprise AI tooling is expected to do. Enterprise AI has spent two years moving from chat responses to AI-assisted drafts. ChatGPT Work moves the target to completed deliverables.
What ChatGPT Work Actually Does
ChatGPT Work consolidates three of OpenAI’s previous agent experiments into a single product:
Operator brought browser interaction: clicking, scrolling, and form-filling on websites and authenticated apps.
Deep Research brought synthesis: pulling structured insight from large bodies of text and data sources.
ChatGPT brought conversational intelligence and task understanding at scale.
Each had a hard ceiling. Operator could not write a detailed research report. Deep Research could not interact with live apps requiring authentication. ChatGPT could draft fluently but could not take action across tools. ChatGPT Work removes those ceilings by combining all three: it can interact with the web, pull from internal business systems, analyze data across sources, and produce complete outputs without a human orchestrating the handoffs between tools.
The product also retires two standalone OpenAI applications. The Atlas browser, which was Operator’s dedicated interface, ends support on August 9. The Codex desktop app is merged into the new unified ChatGPT desktop application for macOS and Windows, available at launch.
For enterprise teams, the practical difference is this: a knowledge worker who currently opens four apps to research a prospect, cross-references CRM records, and then assembles a slide briefing can now describe that outcome once. ChatGPT Work connects to those four apps, runs the steps, and delivers the finished briefing.
The GPT-5.6 Foundation
ChatGPT Work runs on GPT-5.6 Sol, the flagship tier in the GPT-5.6 model family that OpenAI simultaneously opened to the public on July 9. Sol is built for hard problems and long-horizon agentic work. Terra is a balanced everyday model at roughly half Sol’s per-token cost. Luna is the fastest and most affordable option for high-volume simple tasks.
By OpenAI’s own benchmarks, Sol leads competing frontier models on the Coding Agent Index with a score of 80 versus 77.2 for Claude Fable 5, using under half the output tokens. For ChatGPT Work’s long-running workflows, that token efficiency matters: fewer tokens consumed per task step means more steps completed within a given cost envelope. An agent working across a Salesforce data pull, a LinkedIn company research sweep, and a slide generation task burns tokens throughout every step. Sol’s efficiency advantage compounds at scale.
The GPT-5.6 public rollout itself was notable. OpenAI initially shipped Sol, Terra, and Luna to roughly 20 partner organizations at the U.S. government’s request, pending a safety review. That review cleared, and general availability launched July 9 to all customers.
Integration Ecosystem at Launch
The breadth of the launch directory is what makes ChatGPT Work operationally useful immediately rather than aspirationally useful:
| Category | Integrations at Launch |
|---|---|
| File storage | Google Drive, SharePoint, Dropbox |
| Communication and meetings | Slack, Microsoft Teams, Gmail, Outlook, Zoom |
| CRM and revenue | Salesforce |
| Design and creative | Adobe, Canva |
| Code and repositories | GitHub |
| Professional data |
Codex, now integrated into the unified desktop app, also gained 66 single-app plugins at launch, adding Databricks, Hex, Clay, and other data and workflow platforms. Enterprise and Edu admins can deploy workspace-specific app templates for GitHub Enterprise, Snowflake, and Databricks, enabling team-level agent configurations that reflect how a specific department actually works.
The integration count matters because an autonomous agent is bounded by what it can authenticate into. A multi-step workflow crossing three systems requires three live connections. Gaps in coverage create human handoff points that undercut the autonomous premise. The launch directory is competitive with Microsoft Copilot’s connector catalog and substantially wider than what Operator supported at any point during its run.
Enterprise Governance: What Admins Control
When an agent can write to Salesforce, send Slack messages, and commit to GitHub repositories, the error surface expands beyond what a chatbot generates. A misunderstood instruction in a passive system produces a bad draft. A misunderstood instruction in an active agent can modify production records.
OpenAI built ChatGPT Work on top of ChatGPT Enterprise’s compliance foundation and added agent-specific governance controls:
- Workspace admins define which users can launch agents at all.
- Per-agent configurations specify which connected tools are in scope.
- Admins set action permissions at the app level: read-only versus write access per integration.
- A new Global Admin Console surfaces workspace agents with run counts, app permission states, and usage analytics broken down by team and department.
The compliance layer includes SOC 2 Type II certification, end-to-end encryption, data residency options, SSO/SAML, SCIM provisioning for automated user management, IP allowlists, and role-based access control. OpenAI also launched an updated Enterprise Compliance Logs Platform with Admin Audit, User Authentication, and Codex Usage log types at minutes-level latency. For regulated industries, these logs export as immutable JSONL files and integrate with eight DLP and eDiscovery partner platforms.
The least-privilege model, where an agent only accesses what it is explicitly authorized to touch, is the correct pattern for enterprise agent governance. OpenAI has implemented it at the workspace-admin level. This does not make agent deployment consequence-free, but it gives security and compliance teams the controls they need to set boundaries before granting write access to production systems.
What This Means for Enterprise AI Teams
The significance of ChatGPT Work is not any single capability. It is the bundling moment.
Enterprise AI teams have spent 18 months evaluating, piloting, and point-deploying individual AI capabilities in parallel: a research tool in one product, a coding assistant in another, a summarization layer in a third. ChatGPT Work is OpenAI’s answer to integration fatigue: one surface that handles the full workflow rather than one step of the workflow.
As enterprises have struggled to close the gap between AI adoption and AI readiness, the friction has not primarily been model capability. It has been integration complexity, governance requirements, and organizational reluctance to give an AI system write access to production data. ChatGPT Work addresses the first two directly. The third is a change-management problem, and no product launch resolves it.
For teams already on ChatGPT Enterprise, the unlock is immediate: the infrastructure is already approved, the compliance contracts are signed, and adding agent access is an admin configuration decision. For teams evaluating a consolidation of their enterprise AI stack, this is the most complete single-vendor enterprise agent product available as of July 2026.
The competitive context matters. Claude Sonnet 5’s 1M-token context window and agentic tooling push in the same direction. Microsoft Copilot operates across the M365 graph with deep SharePoint and Teams integration. Google’s Gemini Enterprise Agent Platform, previewed at Google I/O in May and delayed to July 17 for an architectural rebuild, targets the same enterprise workflow automation market. The trajectory across all three platforms is consistent: from generative response to completed work product.
OpenAI shipping ChatGPT Work alongside the public GPT-5.6 rollout signals where the product hierarchy goes next. The underlying model is broadly available. The agent wrapper is what enterprise customers are paying for. The integration directory is what makes the agent useful today. The question for enterprise AI leaders is not whether autonomous workflow agents are coming. It is which platform their teams will run them on, and whether they have the governance controls in place to expand agent permissions confidently once the first pilots demonstrate value.
This post reflects publicly available information from OpenAI as of July 11, 2026. Pricing and plan availability may change after publication.