OpenAI Restores GPT-4o After GPT-5 Launch Faces Backlash Over ‘Dumbness’ and Benchmark Blunder

When OpenAI unveiled its much-anticipated GPT-5 model earlier this week, the company framed it as its smartest, most capable AI system to date. But in a turn of events that’s becoming increasingly common in the fast-moving AI industry, the launch was quickly overshadowed by frustrated users, technical hiccups, and an embarrassing “mega chart screw-up” that sparked waves of online ridicule.

Now, in a rare move, OpenAI CEO Sam Altman has announced that GPT-4o—the previous flagship model—will be reinstated for ChatGPT Plus users after widespread complaints that GPT-5 actually felt less intelligent.

From AI Apex to Apparent Step Backward

GPT-5 was pitched as a “unified” model—an ambitious attempt to merge the capabilities of all previous GPT iterations into a single, more seamless system. A central innovation was its real-time router, a mechanism designed to determine on the fly whether a query required a quick, lightweight response or a deeper, more complex reasoning process.

In theory, this would allow GPT-5 to dynamically balance speed and quality. In practice, many users felt the opposite happened. Within hours of launch, Reddit threads and social media timelines filled with gripes about bizarrely shallow answers, missing context, and a general decline in quality compared to GPT-4o.

One Redditor summed it up bluntly: “Feels like I downgraded my subscription.”

Altman addressed these concerns head-on during a Reddit Ask Me Anything (AMA) session Friday:

“GPT-5 will seem smarter starting today. Yesterday, we had a sev, and the autoswitcher was out of commission for a chunk of the day, and the result was GPT-5 seemed way dumber.”

He acknowledged that the real-time router had malfunctioned, causing GPT-5 to repeatedly default to less capable reasoning modes. OpenAI has since patched the issue and is “making interventions to how the decision boundary works” so that the most appropriate internal model is selected more often.

Bringing Back GPT-4o—For Now

OpenAI’s decision to bring back GPT-4o is significant. In the past, the company has phased out older models relatively quickly once a new flagship was released. The return of 4o suggests OpenAI recognizes that its latest model isn’t yet winning over its most engaged users.

“We’re listening,” Altman said. “We will launch 4o again for ChatGPT Plus users and closely review usage before determining how long to support it.”

For now, it appears GPT-4o and GPT-5 will coexist, giving users a choice while OpenAI tweaks its routing system and decision boundaries.

The “Mega Chart Screw-Up” That Went Viral

If performance complaints weren’t enough, OpenAI also had to contend with a presentation mishap that quickly became meme fodder. During GPT-5’s live launch event, a slide comparing benchmark scores appeared to dramatically overstate GPT-5’s accuracy while downplaying that of GPT-4o.

The chart in question related to SWE-bench Verified, a dataset used to measure how well AI systems can fix software bugs. According to OpenAI’s own documentation, GPT-5 achieved 74.9% accuracy with “thinking time” and 52.8% without, compared to OpenAI o3 at 69.1% and GPT-4o at 30.8%. But the visual shown on stage appeared to misalign or swap bar heights, making GPT-5’s lead look far larger than it actually was.

Critics online immediately dubbed it a “chart crime,” and the gaffe prompted a wave of parody graphs on X and Reddit. Altman later admitted on social media that it was indeed a “mega chart screw-up”, though he insisted the official blog post had the correct data.

Still, the incident has raised questions about transparency in AI performance claims—especially at a time when hype cycles are already intense.

A Unified Model—With Growing Pains

The idea behind GPT-5’s unified design is ambitious: instead of forcing users to choose between models optimized for speed (like GPT-3.5 Turbo) or deep reasoning (like GPT-4o), the system would automatically select the right “sub-model” in real time.

Done well, this would mean faster average responses without sacrificing quality where it matters. Done poorly, it means users get a fast answer that isn’t the one they actually wanted. The malfunctioning router appears to have caused exactly the latter.

Altman promised that upcoming updates will make the system more transparent by showing users which model handled their query. That kind of visibility could help rebuild trust, as many power users are accustomed to tailoring their prompts to specific models’ quirks.

Why the Backlash Matters

OpenAI’s audience is uniquely sensitive to changes in quality. Many of the most active ChatGPT Plus subscribers are developers, researchers, and content creators who rely on predictable AI behavior. Even small shifts in reasoning quality or factual recall can break workflows—or worse, undermine trust.

Rolling out a model that feels inconsistent, even temporarily, risks alienating the very people most likely to evangelize it. By bringing GPT-4o back, OpenAI is signaling it values that community’s feedback enough to adapt its rollout strategy in real time.

The Bigger Picture: AI Model Launches Are Getting Riskier

The GPT-5 saga also points to a broader trend: as AI models grow more complex, so do the risks of rolling them out at scale. A single malfunction in something like a routing system can degrade the experience for millions of users simultaneously.

And the “chart crime” episode underlines another risk—perception. In an industry where most users can’t directly verify benchmarks, a single misleading visual can quickly erode credibility, no matter the underlying truth.

What Happens Next

OpenAI says it will:

• Continue supporting GPT-4o for Plus subscribers while refining GPT-5’s routing and decision boundaries.

• Increase transparency about which model is responding to a given prompt.

• Improve internal QA processes to avoid public data presentation mistakes.

In the meantime, user choice has returned—at least temporarily. Those who prefer GPT-4o’s reasoning style can stick with it, while others can test GPT-5 as it evolves.

Altman remains optimistic:

“We think GPT-5 has enormous potential. But it has to earn that trust. That’s what we’re working on now.”

Whether GPT-5 will win back its doubters may depend less on benchmark charts and more on the everyday quality of its answers—the very thing that drew users to ChatGPT in the first place.