ChatGPT Work Cloud adds internet-enabled code execution and persistent filesystem for $20/month users

Work Cloud separates task completion from standard Chat by exposing Luna, Terra, and Sol models at reasoning levels up to Ultra. These options match API tiers yet bill against Codex allowances rather than separate Chat quotas. The code interpreter now permits package installs and external API calls that remain blocked in Chat sessions, extending the 2023 Code Interpreter pattern beyond the restricted allowlists Claude implemented in 2025.

Simon Willison documented the split between Work Cloud and Work Local but did not address model naming opacity. Sol, Luna, and Terra lack public architecture papers or benchmark tables, unlike GPT-4o or o1 releases. This absence reduces user ability to predict cost or output variance when selecting reasoning depth, undermining the transparency required for sustained enterprise trust.

Persistent filesystem and scheduled automations create stateful workflows absent from stateless Chat. Combined with publishable Sites, these features shift usage from one-off queries toward repeatable pipelines. Operational impact appears first in data analysis and reporting tasks where prior sessions retain intermediate artifacts.

Future iterations will likely expose the same container controls through the API, aligning ChatGPT Work with existing Codex billing. Adoption metrics will hinge on whether Ultra mode delegation exceeds the 30 percent sub-agent rate observed in early Codex logs.