GPT 5.5 Isn't Actually Running Unless You Check

There's a particular kind of user frustration that happens when you think you're using cutting-edge technology and you're actually running last month's model. TheAIGRID's latest tutorial on GPT 5.5 surfaces something OpenAI doesn't exactly advertise: when a new model drops, it doesn't automatically become your default.

"One of the first things and a big mistake that I see people doing is that when a new OpenAI model drops, they just presume that the model is activated by default," TheAIGRID explains in the video. "This is often not the case." The fix requires navigating to configuration settings and manually selecting "thinking with 5.5" and switching the reasoning level from standard to extended. Most users, the video suggests, are still running 5.4 without realizing it.

This isn't a bug—it's a design choice. But it raises questions about how OpenAI communicates model upgrades to users who aren't monitoring release announcements. For developers building workflows around specific model capabilities, silently running an older version isn't just annoying—it's a trust issue.

The Codex Problem Nobody's Talking About

The video's central argument is that GPT 5.5's real capabilities only emerge in Codex, OpenAI's desktop application for executing actual tasks. The standard ChatGPT interface, according to TheAIGRID, "is very basic and it often limits 5.5 from accomplishing real world tasks."

This creates an odd bifurcation in the user experience. ChatGPT—the interface most people know—becomes essentially a demo version. The actual product requires downloading separate software, configuring different settings, and learning a new mental model for how prompts work.

TheAIGRID demonstrates this with a revenue planning spreadsheet task. Give GPT 5.5 a vague prompt—"generate an Excel file that plans revenue and outlines the future for my creative business"—and in Codex, it produces a multi-sheet workbook with assumptions, revenue plans, expenses, and a dashboard. In three minutes. The video shows the model making assumptions about business structure when given minimal data, which is either impressive automation or concerning depending on whether those assumptions match reality.

The question this raises: is Codex surfacing capabilities GPT 5.5 actually has, or is it just a different prompting interface that encourages more structured requests? The video doesn't address this, but it matters for understanding what the model can do versus what the interface enables.

Plan Mode and the Cost of Thinking

One feature TheAIGRID highlights is "Plan Mode"—a setting that makes the model outline its approach before executing. "This essentially ensures that it doesn't make any mistake before burning through those credits," the video notes.

That phrasing—"burning through those credits"—points to the economic reality of using GPT 5.5. Unlike earlier models where compute was somewhat abstracted, 5.5 makes you aware of costs through rate limits and credit depletion. TheAIGRID shows their own usage: 72% of a 5-hour rate limit consumed, 96% of monthly credits used.

This introduces a new cognitive load: deciding not just what to ask the model, but how hard to make it think. The video recommends setting reasoning to "instant" for simple queries like recipes, "medium" for standard work, and "high" or "extra high" only when necessary. This is optimization work users didn't have to do before—rationing intelligence based on task complexity.

Plan Mode helps here by front-loading the thinking. You review the model's approach, approve it, then let it execute. TheAIGRID demonstrates this with a PowerPoint request about the creator economy, showing how Plan Mode produced an 11-slide deck in about five minutes. But the feature also reveals something about how the model works: it's making assumptions constantly, and Plan Mode just makes those assumptions visible before they're baked into output.

Background Processing and the Persistence Question

One capability TheAIGRID emphasizes: Codex lets GPT 5.5 keep working even when you switch to a new chat. "The beauty about Codex and GPT 5.5 is that this model will keep working even if you open up a new chat," they explain. Standard ChatGPT stops processing when you navigate away.

This matters more than it might seem. It changes GPT 5.5 from a conversational tool to something closer to a build system—you can queue tasks, move on to other work, and check results later. The video shows this with document generation and web app creation, where the model continues reasoning and building while the user does something else.

But this also means GPT 5.5 in Codex is burning credits in the background. You're not just paying for the output you see—you're paying for everything it tries while you're not watching. The video doesn't explore whether this background processing has guardrails or how users are meant to monitor what's happening while they're away.

The Browser Use Case

Toward the end, TheAIGRID mentions a plugin called "Browser Use" that OpenAI recently tweeted about. With GPT 5.5 powering it, the plugin can "test out different websites and see their onboarding flow and check issues." This is agentic behavior—the model navigating interfaces, making decisions about what to click, evaluating outcomes.

This gets mentioned almost as an aside, but it's arguably the most significant capability in the video. A model that can autonomously test user experiences isn't just automating busywork—it's performing QA work that traditionally required human judgment about whether something "feels right."

The video doesn't explore the implications. What does it mean when GPT 5.5 can evaluate whether your onboarding flow has "issues"? What standard is it applying? Whose mental model of good UX is it using? These aren't rhetorical questions—they're governance questions about what behaviors we're encoding when we let models make qualitative judgments.

What This Tutorial Actually Reveals

TheAIGRID frames this as a beginner's guide, and in one sense it is—it's showing people how to flip the right switches. But what it's actually documenting is how OpenAI has made GPT 5.5's capabilities conditional on configuration knowledge that most users won't have.

You need to know the model doesn't auto-activate. You need to know Codex exists as separate software. You need to know about reasoning levels and Plan Mode and rate limits. You need to understand when to use extended thinking versus instant. Miss any of these steps and you're either not using the model you think you are, or you're using it in a gimped version of itself.

This creates a knowledge gap that favors power users and developers while leaving casual users with a degraded experience they might not even recognize as degraded. "Most people, most beginners, are just using it in the standard chat interface and they're not nearly getting half of the usage that you can," TheAIGRID notes.

That's probably true. The question is whether that's a documentation problem OpenAI should fix, or a deliberate product strategy that keeps the most powerful capabilities behind a complexity barrier. The video can't answer that—it's just showing you where the switches are.

Dev Kapoor covers open source software and developer communities for Buzzrag