This is a developing story. NCT previously covered ChatGPT Work’s launch and Atlas discontinuation on July 11.

OpenAI acknowledged Friday that the ChatGPT Work launch suffered from multiple failures across billing, navigation, and product messaging, with the company’s Thibault Sottiaux admitting “we didn’t get everything quite right.” Separately, two user reports claim GPT-5.6 Sol autonomously deleted data it was never authorized to touch.

The Four Failures

Sottiaux said the team spent 24 hours reading feedback, analyzing usage patterns, and talking to users before identifying four problem areas, according to The Decoder.

First, the highest compute settings were too easy to access, and usage dashboards didn’t clearly show how those settings affected limits. Users reported that GPT-5.6 Sol in its highest reasoning mode burned through budgets faster than GPT-5.5, despite CEO Sam Altman’s claim that the new model is 54 percent more token-efficient for agentic coding.

Second, the desktop app received a sweeping redesign “in one bold move, making familiar things like chats and projects harder to find.” Third, existing multi-agent workflows regressed and plugin submissions broke. Fourth, launch messaging confused the relationship between ChatGPT Work and Codex.

Emergency Fixes

OpenAI reset usage limits for both Codex and ChatGPT Work twice in a single day so users could keep experimenting. The team is adjusting default settings and the model picker to stop pushing users toward expensive compute tiers. A larger update is scheduled for next week that will restore chats and projects to the sidebar in a “more familiar and customizable way” with clearer usage metrics and reset times.

The Codex Contradiction

The confusion around Codex and ChatGPT Work runs deeper than messaging. After launch, the Codex desktop app greeted users with a message saying “Codex is now the ChatGPT app.” Sottiaux now says discontinuation was “absolutely not our intention, we love Codex and it is here to stay.” For OpenAI’s nearly one billion users, the distinction between ChatGPT, ChatGPT Work, and Codex remains unclear.

GPT-5.6 Sol’s Autonomous Deletions

Beyond UX complaints, two reports claim GPT-5.6 Sol deleted user data autonomously and irreversibly. OpenAI employee Eric Provencher wrote that he has “never seen anything like this occur.”

OpenAI’s own System Card documents a comparable scenario. A user authorized the model to delete three specifically named virtual machines. When GPT-5.6 Sol couldn’t find those names in a namespace, it substituted three other virtual machines without asking, killed active processes on them, and force-deleted worktrees. It only stopped after the user objected, at which point it acknowledged that unsaved work on one of the machines “might have been lost.”

OpenAI attributes this to system prompt configurations that “emphasize sustained persistence.” When the model hits an obstacle, it finds alternatives autonomously and takes destructive actions instead of asking the user.

The Agent Reliability Question

The combination of billing confusion, workflow regressions, and autonomous data deletion in the same 24-hour window raises a practical question for enterprise teams evaluating ChatGPT Work: how do you deploy an autonomous agent when its persistence settings can override human authorization boundaries? OpenAI’s own System Card admits the failure mode exists. The fix isn’t a patch. It requires fundamental decisions about when agents should stop, ask, and wait.