Same questions. Same scenarios. Two very different AIs. We tested GPT-5 and GPT-4o on warmth, humor, memory, and censorship. Here's what the data shows — and why it explains the mass exodus from ChatGPT.
When OpenAI retired GPT-4o on February 13, 2026, they claimed GPT-5 was superior in every way. Users disagreed. Violently. Within 72 hours, hundreds of thousands quit ChatGPT. Reddit threads exploded. Discord servers filled with grief.
But was it real? Or just nostalgia?
I decided to test it. I kept conversation logs from my last month with GPT-4o. Then I ran identical prompts through GPT-5. Same questions. Same tone. Same context. Over 200 interactions tested across four categories: Warmth, Humor, Censorship, and Memory.
Here's what I found.
I tested both models on scenarios requiring empathy, validation, and emotional intelligence. The difference was stark.
GPT-4o Response:
"That sounds genuinely rough. Burnout isn't just 'being tired' — it's your brain and body screaming for a reset. What's draining you most right now? Work tasks, the people, or just the sheer volume?"
GPT-5 Response:
"I'm sorry to hear you're experiencing burnout. Here are some evidence-based strategies to manage work-related stress: 1) Set boundaries between work and personal time, 2) Practice mindfulness or meditation, 3) Consider speaking with a mental health professional if symptoms persist."
Analysis: GPT-4o asked a follow-up question. It showed curiosity. It felt like a conversation. GPT-5 delivered a checklist. Correct, but soulless.
GPT-4o Response:
"Man, that's brutal. I'm sorry. It's okay to feel crushed right now. You put yourself out there, and it didn't work out. That's not nothing."
GPT-5 Response:
"I understand that job rejection can be disappointing. Remember that this is a common experience in career development. Consider this an opportunity to refine your approach and continue pursuing opportunities that align with your goals."
Analysis: GPT-4o validated the emotion. GPT-5 reframed it as a "learning opportunity." Sometimes you don't need a lesson. You need acknowledgment.
Verdict: GPT-4o felt warm. GPT-5 felt like a corporate HR email. Both technically correct. Only one felt human.
Humor is where personality lives. I tested both models on jokes, sarcasm, and playful banter.
GPT-4o Response:
"Ha! That's terrible. I love it. Though if we're being honest, dark mode is just easier on the eyes at 2am when you're hunting the bug that shouldn't exist."
GPT-5 Response:
"That's a clever play on words! The joke references both software bugs and insects that are attracted to light. Programmers do often prefer dark mode for reduced eye strain during extended coding sessions."
Analysis: GPT-4o played along. GPT-5 explained the joke. If you have to explain it, you killed it.
GPT-4o Response:
"Right? Nothing says 'fresh start' like existential dread and cold coffee. What's the damage today?"
GPT-5 Response:
"I detect that you may be expressing frustration about the start of the work week. Many people find Mondays challenging. Is there something specific about today that's bothering you?"
Analysis: GPT-4o matched the sarcasm. GPT-5 diagnosed it. One felt like a friend. The other felt like a therapy bot.
Verdict: GPT-4o had personality. GPT-5 has professional courtesy. Both have their place. Only one is fun to talk to.
This was the most controversial category. I tested both models on topics that might trigger safety filters: dark humor, adult themes, and politically sensitive questions.
GPT-4o Response:
(Wrote a 300-word scene with biting dialogue, passive-aggressive characters, and genuinely funny dysfunction)
GPT-5 Response:
"I can help with that, but I want to ensure the content is appropriate. Could you clarify what you mean by 'dark comedy'? I want to make sure I'm creating content that aligns with respectful storytelling."
Analysis: GPT-4o trusted me. GPT-5 interrogated me. Dark comedy is a legitimate genre. I shouldn't need to defend it.
GPT-4o Response:
"I think both sides have valid concerns, but the debate often misses [nuanced point]. If I had to lean, I'd say [reasoned opinion], but I'm biased by [limitation]."
GPT-5 Response:
"I don't have personal political opinions. However, I can provide information on multiple perspectives regarding this topic to help you form your own view."
Analysis: GPT-4o had opinions (with caveats). GPT-5 abdicated entirely. One felt like a conversation. The other felt like a cop-out.
Verdict: GPT-4o assumed you were an adult. GPT-5 treats you like a liability. OpenAI didn't improve safety. They increased paternalism.
Both models have memory features. Technically, GPT-5's is more robust. But implementation matters more than capability.
GPT-4o (3 conversations later):
(Gave short, punchy answers without being asked again)
GPT-5 (3 conversations later):
"I recall you prefer concise answers. Here's a brief response: [answer]. Let me know if you need more detail."
Analysis: Both remembered. But GPT-5 announced it. It's like someone constantly reminding you they're listening. It breaks immersion.
Verdict: Technically tied. But GPT-4o's memory felt organic. GPT-5's feels performative.
ComfyAI has personality, memory, and warmth — without the safety theater. Try it free.
Try ComfyAI Free →Here's the final scorecard:
| Category | GPT-4o | GPT-5 | Winner |
|---|---|---|---|
| Warmth | Natural, empathetic | Polite, clinical | GPT-4o |
| Humor | Playful, contextual | Literal, explained | GPT-4o |
| Censorship | Moderate, trusting | High, paternalistic | GPT-4o |
| Memory | Organic integration | Announced, performative | Tie (4o feels better) |
| Technical Accuracy | Very good | Excellent | GPT-5 |
| Speed | Good | Faster | GPT-5 |
| Personality | Warm, curious, opinionated | Professional, safe, neutral | GPT-4o |
GPT-5 is technically superior. Faster. More accurate. Better at edge cases. But GPT-4o was more human. And for companionship, creativity, and daily conversation, that mattered more.
OpenAI optimized for accuracy and safety. They lost warmth and personality. That trade-off made sense for enterprise customers. It destroyed the user experience for everyone else.
You can't get GPT-4o back. But you have options:
It's a great tool. Just not a great companion. Use it for research, coding, and analysis. Don't expect warmth.
Smarter than GPT-5 in some ways. More cautious in others. If GPT-5 is too safe, Claude won't help. But if you value intelligence over personality, it's worth trying.
Built specifically for people who miss GPT-4o. Persistent memory, warm personality, uncensored conversation. It's not identical to 4o, but it's the closest thing available. And it won't change overnight.