OpenAI’s $20 ChatGPT Work Targets Every Office, but Trust and Costs Loom
OpenAI's $20 ChatGPT Work brings autonomous agents to office tasks, but weak adoption, broad permissions, high token costs and rivals test its appeal.
Summary
OpenAI’s July 2026 launch of ChatGPT Work brings autonomous, multistep agents to non-engineers for $20 monthly, its lowest subscription tier. Derived from Codex, it connects with email, browsers, Slack, Notion, Figma, Salesforce, calendars and other SaaS tools to create reports, dashboards, spreadsheets, investment memos and visualizations. Andrew Ambrosino, OpenAI’s desktop app lead, grants it extensive access despite possible private-message leakage. An OpenAI-backed June study found Codex use among 98% of employees, 17% of organizational subscribers and under 1% of individual subscribers. Work and Codex jointly have 20 million users, versus more than 1 billion for ChatGPT online.
Permissions remain confusing, full access may be required where read-only access fails, key settings are web-only, Google Calendar integration cannot create calendars, and low effort settings produce weak results. OpenAI evaluates performance through GDPval, covering 44 occupations and hundreds of knowledge-work tests, plus feedback and optional training traces. Akshay Nathan says reducing complexity remains central, while Joe Gershenson acknowledges reasoning controls need improvement.
OpenAI is chasing Anthropic’s Claude Cowork and Claude Code, Perplexity’s browsing agent, specialist platforms Harvey and Clay, and model-agnostic or open-source alternatives. Codex overtook Claude Code downloads in April, but Databricks found Earendil’s Pi harness outperforming Codex with the same GPT 5.5 model. Mario Zechner argues office outcomes are harder to capture than coding traces and warns proprietary harnesses create lock-in. One four-day test consumed over 80 million tokens, estimated at $65, more than three times the monthly price. Thibault Sottiaux cited an 80% Luna price cut and expects comparable tasks to cost less within six months.
Positives
- ChatGPT Work costs $20 monthly and extends Codex-style autonomous, multistep work to OpenAI’s lowest subscription tier.
- 98% of OpenAI employees used Codex in June, demonstrating strong internal adoption of agentic workflows.
- ChatGPT Work can generate recurring reports, dashboards, spreadsheets, calendars, databases, investment memos and research updates.
- Codex moved slightly ahead of Claude Code downloads after April, with enterprise surveys also indicating OpenAI is catching up.
- OpenAI cut Luna model pricing by 80%, while Thibault Sottiaux expects further efficiency gains within six months.
Risks & concerns
- Under 1% of individual subscribers and 17% of organizational subscribers used Codex in June, despite 98% adoption inside OpenAI.
- ChatGPT Work and Codex have 20 million users, far below ChatGPT’s claimed online audience of more than 1 billion.
- Broad access to inboxes, private messages and workplace tools creates disclosure, privacy and data-retention risks.
- Read-only permissions produced errors, some connections demanded full access, and important controls were limited to the web app.
- More than 80 million tokens consumed in four days cost an estimated $65, exceeding three times the $20 monthly subscription price.
- Databricks found the open-source Pi harness beat Codex using the same GPT 5.5 model, challenging OpenAI’s product advantage.