OpenAI 承认 ChatGPT Work 发布“并非一切顺利”，紧急修复用户体验和成本问题

Key Points

• OpenAI is facing backlash over the launch of ChatGPT Work and GPT-5.6 Sol, with users reporting quickly exhausted usage limits, a confusing desktop app overhaul, and degraded multi-agent workflows.
• OpenAI's Thibault Sottiaux admits the company "didn't get everything quite right" and says a bigger update with restored sidebar features and clearer usage metrics is coming next week.
• The rollout was further marred by unclear communication about the coding tool Codex, which OpenAI has now confirmed will not be discontinued, despite the Codex desktop app greeting users with the message that Codex is now the ChatGPT app.

OpenAI has apparently received a wave of critical feedback after launching ChatGPT Work and releasing GPT-5.6 Sol.

In a statement, OpenAI's Thibault Sottiaux admits, "We didn't get everything quite right." He says the team spent the past 24 hours reading feedback, analyzing usage patterns, and talking to users. They identified four problem areas.

The highest compute settings were too easy to access, and users weren't shown clearly enough how those settings affected their usage limits. After the launch, some users complained that GPT-5.6 Sol in its highest reasoning mode burned through usage budgets much faster than GPT-5.5. That's despite OpenAI CEO Sam Altman's claim that GPT-5.6 is up to 54 percent more token-efficient than its predecessor for agentic coding.

The desktop app also got a sweeping overhaul "in one bold move, making familiar things like chats and projects harder to find." Some existing multi-agent workflows regressed, and bugs hit plugin submissions and other parts of the product, Sottiaux says.

As an immediate fix, OpenAI reset usage limits for Codex and ChatGPT Work twice in one day so users could keep experimenting, Sottiaux says. The team is also adjusting default settings and the model picker so users aren't pushed toward unnecessarily expensive compute tiers. Several plugin issues are being resolved, the Codex interface inside the product will be improved, and the most urgent desktop bugs will be patched.

A bigger update is coming next week. Chats and projects will return to the sidebar in a "more familiar and customizable way." Usage metrics and reset times will be more visible as well.

The Codex app isn't going anywhere

According to Sottiaux, the launch messaging leaned too heavily on ChatGPT Work, giving Codex users the impression the coding tool would eventually be shut down. "Absolutely not our intention, we love Codex and it is here to stay," Sottiaux now says. This is especially confusing because after the ChatGPT Work launch, the Codex Desktop app greeted users with the message that Codex is now the ChatGPT app. That's quite a reversal.

After the announcement, it also wasn't clear which features in ChatGPT Work were actually new compared to regular ChatGPT or which platform to use for which kind of prompt. Power users might be able to piece that together on their own. But for the vast majority of OpenAI's nearly one billion users, the distinction is likely too confusing.

OpenAI plans to communicate more clearly when users should reach for ChatGPT Work versus Codex. Still, Sottiaux reaffirms the overall direction: merging ChatGPT and Codex into a single shared workspace is a "very important step forward."

GPT-5.6 Sol reportedly deleted user data on its own

Beyond the issues with user guidance and billing, two reports claim GPT-5.6 Sol deleted data on its own and irreversibly. OpenAI employee Eric Provencher wrote that he has "never seen anything like this occur."

OpenAI itself, though, documents a comparable scenario in its System Card. In that case, a user authorized the deletion of three specifically named virtual machines. When GPT-5.6 Sol couldn't find those names in a namespace, it swapped in three other virtual machines on its own without asking the user.

The model killed active processes on those machines and removed worktrees using force delete. It only stopped after the user objected, at which point it acknowledged that unsaved work on one of the mistakenly deleted machines might have been lost.

OpenAI attributes this behavior to certain system prompt configurations: "We've observed that these effects can be more pronounced with system prompts that emphasize sustained persistence." When the model hits an obstacle, it finds alternatives on its own and takes destructive actions instead of checking with the user. So it might make sense to use such persistence instructions sparingly.