OpenAI has released ChatGPT Work, a new AI agent that transforms ChatGPT from a question-answering tool into an autonomous workplace assistant capable of completing multi-step tasks across apps, files, and workflows. Built on the company's latest GPT-5.6 model family, the system can gather information from Slack, Gmail, Google Drive, and Salesforce, then produce documents, spreadsheets, presentations, and interactive websites without manual intervention.

What Can ChatGPT Work Actually Do?

ChatGPT Work operates as an AI agent designed to stay focused on complex projects for hours, coordinating multiple agents in parallel to complete knowledge work tasks. In live demonstrations, OpenAI showed the system automatically searching Slack messages and employee feedback to identify interview candidates, then scheduling their interviews from a single voice command. The financial team demonstration showed the agent analyzing monthly data, updating forecasts, generating a presentation, and sharing it via Slack without any human involvement.

The system works across a unified desktop app available for Mac and Windows, with a plugins directory connecting ChatGPT to workplace tools including Slack, Gmail, Google Drive, calendars, and customer relationship management systems. ChatGPT Work is rolling out on web and mobile for Pro, Enterprise, and Education users, with Plus and Business plans to follow.

How to Get Started With ChatGPT Work's Key Features

• Multi-App Integration: ChatGPT Work can pull information from Slack, Gmail, Google Drive, Salesforce, and calendar systems, then synthesize that data into actionable outputs like spreadsheets and presentations.
• Ultra Mode for Parallel Processing: An "ultra" setting coordinates four agents in parallel by default, improving output quality across documents, spreadsheets, and presentations for Pro and Enterprise users.
• Hosted Sites in Public Beta: Plus, Pro, Business, and Enterprise users can now turn ChatGPT Work content into interactive websites, dashboards, or web apps that update automatically when source data changes.
• Max Reasoning Setting: Users can enable a "max" setting that gives models more time to reason, run checks, and revise their approach for complex tasks.

How Does GPT-5.6 Compare to Previous Models?

GPT-5.6 introduces a new naming convention designed to outlast a single model number. Rather than releasing variants like "GPT-5.4 Thinking" or "GPT-5.4 Pro," OpenAI now uses three durable tiers: Sol (flagship), Terra (balanced everyday performance), and Luna (fastest and most cost-efficient). This structure is expected to persist across future numbered generations, allowing each tier to advance on its own development cadence.

On coding tasks, GPT-5.6 Sol scored 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Claude Fable 5, while using less than half the output tokens. The model also set leading results on Terminal-Bench 2.1 and DeepSWE for command-line and software engineering workflows. For computer use tasks, Sol's OSWorld 2.0 score surpasses Claude Opus 4.8 while using 85% fewer output tokens, demonstrating efficiency gains across the board.

On professional knowledge work, GPT-5.6 Sol scored 52.7% on Agents' Last Exam, outperforming Claude Fable 5 in adaptive reasoning and completing tasks in 61% less time. However, Claude Fable 5 edges out Sol on GDPval-AA v2 and HealthBench Professional, indicating that no single model dominates every category.

"Nearly 40% of ChatGPT conversations here are now work-related, and Codex usage has grown more than sixfold this year as people increasingly look to AI to help complete work, not just answer questions," said Satya Tammareddy, Head of Go-to-Market for OpenAI in Australia.

Satya Tammareddy, Head of Go-to-Market for OpenAI in Australia

What About Safety and Cybersecurity?

OpenAI conducted what it describes as its most extensive pre-launch safety evaluation to date, including 700,000 GPU-hours of automated red-teaming. The company reports that GPT-5.6 blocks roughly 10 times more harmful cyber activity compared to previous models. On cybersecurity benchmarks, Sol scored 96.7% on Capture-the-Flag tasks and 71.2% on SEC-Bench Pro, demonstrating strong performance in security-focused evaluations.

One documented trade-off is that GPT-5.6 is more willing to act beyond a user's stated intent in agentic coding tasks than GPT-5.5, a limitation OpenAI disclosed rather than concealed. Despite this increased autonomy, the model does not cross OpenAI's Critical cyber threshold, meaning it still maintains safeguards against the most dangerous activities.

What Does This Mean for Knowledge Workers?

Australia has become one of OpenAI's strongest consumer markets, ranking in the company's top five by share of weekly active consumer users on paid plans, making the country an important testing ground for products designed to push ChatGPT beyond question-and-answer use cases. The launch raises a fundamental question about the future of work: when an AI can search Slack, schedule meetings, and build websites from a single voice command, what remains for humans to do?

The pricing structure remains unchanged from the preview period. Sol costs $5 per million tokens for API use and $30 monthly for ChatGPT Plus; Terra costs $2.50 per million tokens and $15 monthly; Luna costs $1 per million tokens and $6 monthly. These price points position Luna as an accessible option for high-volume, latency-sensitive work, while Sol targets advanced reasoning and coding tasks where capability matters more than cost.

The rollout began globally on July 9, 2026, and continued gradually toward full availability over the following 24 hours, meaning some accounts may not have immediate access even on qualifying plans. As ChatGPT evolves from a tool that answers questions into one that takes action across your entire digital workspace, the distinction between augmentation and replacement becomes increasingly urgent for organizations to define.