ChatGPT Work

Executive Summary

ChatGPT Work is OpenAI's agent experience for assignments that need multiple steps and a finished deliverable: research, spreadsheets, presentations, documents, and recurring updates. It can retrieve information from approved tools, create artifacts, and accept steering during execution. OpenAI describes Work and Codex as sharing core capabilities, with Work organized around general tasks rather than a software-development interface.^[1]

Its place in the personal-agent landscape depends on the execution path. Cloud tasks run on OpenAI infrastructure and can continue away from the device. Local tasks can work with approved files and applications, but local execution still sends relevant context to hosted model services. Neither the desktop interface nor the word “local” establishes that all data stays on the machine.^[2][3]

Availability remains subject to rollout and workspace policy. Cloud and local access can become available at different times for the same account.^[4]

---

Product Overview

The user chooses Work, supplies sources, and describes an outcome. A project keeps related context together; plugins add connected systems; the result can be refined in the same conversation. The web guide explicitly includes documents, spreadsheets, presentations, PDFs, and research workflows.^[5]

Key Capabilities

Product Surfaces

The desktop documentation provides macOS and Windows downloads plus a Linux installation guide. This does not establish identical tools on all three: native Computer Use is documented for macOS and Windows.^[6][9][4]

A practical first task is a source-backed comparison with an editable spreadsheet and explicit review criteria. Ask for uncertainty to be recorded, and specify which actions should wait for review. This is a suggested evaluation workflow, not a benchmark result.

---

Technical Architecture

Cloud execution

OpenAI documents Work Cloud as the Codex harness in VM-backed sandboxes on shared, managed infrastructure. Environments can be reused or replaced while eligible state persists; this is not a promise of a fresh container for every task or a dedicated physical host per customer. Cloud work does not inherit the initiating device's local files, signed-in browser profile, or VPN access.^[2]

Cloud browser access, web search, connected applications, and code/shell networking have separate controls. Turning off general network access for code still permits required destinations on a managed allowlist and does not switch off those other tools. Website permission also differs from approval for a consequential action on that website.^[11]

Local execution and permissions

Files and shell commands follow the task's sandbox and approval settings. Native applications operate with their own account and operating-system permissions. A file restriction on direct tools therefore does not prove that an approved application cannot open that file. Browser Use site policies also do not apply when Computer Use operates a browser as a native application.^[3]

On macOS, Computer Use needs Screen Recording and Accessibility permissions and supports background operation. Windows Computer Use takes over the visible, active desktop; it cannot simultaneously operate in the background of that same user session. These differences matter when choosing a machine for unattended work.^[9]

Model choice

The current guide distinguishes GPT-6 Astra from GPT-5.6 Sol, Terra, and Luna. Reasoning and model selection change resource use as well as the kind of task to evaluate. Astra is available in Work on eligible Plus plans with limited included usage; Pro $100/$200 and Business Premium seats can apply their full existing allowance to it. This is an allowance policy, not an unlimited-use promise.^[12][4]

Persistence, integrations, and recurring work

A plugin's installation, the underlying app's availability, account authorization, and action permissions are separate decisions. Some plugins are desktop-only. Hooks require their scripts to exist in the execution environment; installing a plugin on the web does not deploy a local script there.^[7]

Cloud schedules can run without the user's computer. Local schedules need that computer and app running. Event-triggered tasks currently belong to eligible web/mobile experiences, and one task cannot combine an event trigger with a time schedule. Test the prompt and its tool access before making it recurring.^[8][1]

Memory is useful recall rather than an immutable policy store. OpenAI recommends keeping required team rules in explicit instructions or checked-in documentation. Chat deletion and memory deletion also have different effects; inspect the actual memory controls when removing sensitive or outdated context.^[10][11]

---

Strengths

• One place for sources and artifacts. The desktop workspace can display the files produced during a longer assignment, making review part of the workflow.^[6]
• Cloud continuity. Supported cloud conversations and tasks can continue across web, mobile, and desktop without keeping a local machine awake.^[4][1]
• Reusable connected workflows. Plugins, skills, schedules, and event triggers support repeatable work with explicit source access.^[7][8]
• Adjustable reasoning. Available model and reasoning settings let users change the resource commitment for a task; Ultra can distribute suitable work among subagents.^[12]

---

Cautions

• Local does not mean offline. File excerpts, screenshots, prompts, and tool results can enter OpenAI services even when tools run on the user's computer.^[3]
• Permissions have multiple paths. A shell network restriction does not cover every browser or connector action, and a shared connection may carry different permissions from the requesting user's personal account.^[11]
• Retention is not one switch. Conversations, execution snapshots, Library files, project files, memories, and connector records have distinct lifecycles. Deleting a chat does not instantly purge them all.^[2]
• Usage is workload-dependent. Longer tasks, expensive models, repeated runs, and higher reasoning settings can consume allowance rapidly; messages are not a fixed measure of work.^[13]
• Automation needs maintenance. The model guide schedules GPT-5.5 retirement from ChatGPT, Work, and Codex for October 14, 2026. Saved configurations that pin it need an available replacement; this retirement does not apply to the API.^[12]

---

What Developers Say

On September 12, 2026, evilintala reported on OpenAI's community forum that Work with GPT-5.6 Sol lost context, stalled, and made changes beyond a requested UI fix. The thread was automatically closed after a day, which does not establish that the reported behavior was resolved. This is one user's experience, not a measured failure rate or a finding reproduced for this review.^[14]

A separate September 4 report by h3yshayna describes Linux desktop build 26.901.20858 creating local-only conversations through New chat while existing cloud conversations synchronized. That issue illustrates why platform and conversation mode should be recorded in a pilot; it does not prove that every Linux installation or later build behaves the same way.^[15]

The documentation establishes published capabilities. This review did not run a comparative workload or independently measure output accuracy, completion rate, or total cost per finished deliverable.

---

Pricing & Licensing

Selected published US-dollar plan prices checked September 16, 2026:

Work and Codex share usage rather than providing independent unlimited budgets. Task size, model, context, and tool use affect consumption; cloud tasks can consume more than a simple local interaction.^[13]

For credit-based enterprise agreements, using already committed credits is different from incurring an additional invoice charge. User limits and workspace overage controls are separate; an alert is a notification, not a spending cap. Evaluate actual billing against the agreement and invoice.^[16]

Licensing model: Commercial ChatGPT service. The plan is not a self-hosted Work Cloud license.

Additional costs: Optional credits, permitted overages, connected-service charges, and the time spent reviewing or correcting results. No fixed cost per completed task was established here.

---

Competitive Positioning

In personal-agent platforms, Work fits users who want documents, analysis, connected tools, and repeatable tasks inside an existing ChatGPT account. Its local/cloud choice is consequential: select the execution path that can reach the required sources and remain available for the task.^[1]

Tembo is an adjacent option when the assignment is a team software workflow. Its current platform runs coding agents across repositories, tickets, and tools in cloud VMs, with pull requests or artifacts for review and a self-hosted deployment option. Work's focus here is general personal and business deliverables. Disclosure: Ry Walker is Tembo's founder and CEO.^[17][5]

---

Ideal Customer Profile

Best fit: Individuals and teams already using ChatGPT who can define a deliverable, supply authorized sources, inspect the result, and choose appropriate permissions and usage limits.

Poor fit: Workflows requiring fully offline inference, a customer-operated Work Cloud runtime, or identical application automation across every desktop platform.

---

Viability Assessment

Current documentation covers deployment boundaries, administration, billing, plugins, and recurring work in substantial detail. That supports evaluating a shipped product with specific controls; it does not establish an independent reliability ranking or a separate Work revenue figure. Rollouts, model retirements, and the cited user reports make a representative pilot more useful than a capability checklist alone.^[4][12][14]

---

Bottom Line

ChatGPT Work is a credible choice for substantial tasks whose output a person can inspect. Its useful unit is the completed, reviewed deliverable. Evaluate local and cloud execution separately, and measure recurring consumption alongside quality and the amount of manual correction required.

Recommended for: Source-backed research, document production, analysis, and recurring connected workflows with human review.

Not recommended for: Assuming that local tools guarantee local data processing or that an included subscription permits unlimited unattended work.

Outlook: Strong scope for general work, with practical value dependent on the task, available integrations, permissions, and actual cost of a reviewed result.

---

Research by Ry Walker Research • methodology