Every new SOTA (state of the art) model uses the same announcement template: benchmark graphs, travel planning demos, and claims that this time, the model really understands you.

OpenAI seems to finally recognize those benchmarks don't matter to most users. ChatGPT 5.1's announcement buried the graphs and led with personality switching, adaptive thinking, and conversational warmth.

Translation: they're selling vibes now.

Which makes sense.

When they made GPT-5 the default and discontinued 4o, users revolted. Not because 5 was less capable, but because it felt robotic. 4o had warmth.

5 had efficiency.

So they brought 4o back for plus users.

Now we have 5.1 with six personality presets.

But does it actually fix what users missed about 4o?
Is "conversational AI" just vibes, or does it matter?

I tested it. Here's what I found.

1. Marble: World Labs' 3D world model

AI model that generates explorable 3D environments from text or images. Creates scenes with physics, lighting, and spatial consistency in real-time. Can navigate and modify generated worlds on the fly.

2. Lua AI: Unified agent platform for mid-market enterprises

It reverse-engineers websites' network requests to generate deterministic APIs - no headless browsers needed. Launched in August, hit 30K users in 3 weeks. Claims 10-100x faster than browser automation, costs fractions of a cent per call. Handles dynamic values, IP rotation, antibot challenges automatically.

3. Hillclimb: Training data derived from human superintelligence

Generates AI training data by capturing expert decision-making processes in real-time. Instead of generic synthetic data, records how top performers actually solve complex problems. Claims to create higher-quality training sets by learning from human expertise.

Test 1: The em-dash promise (follows instructions better)

Sam Altman claimed that if you tell ChatGPT not to use em-dashes, it won't. This used to be impossible because models would ignore style instructions and revert to their defaults.

I tested it on a generic writing task, explicitly telling them to avoid em-dashes.

Result: It complied perfectly.

The model actually listens to style instructions now.

A big deal if you've ever tried to make AI writing look less robotic.

Test 2: Six personalities, one answer

OpenAI's pitch: you can now switch between six personalities i.e, Default, Professional, Efficient, Candid, Cynical, and Friendly to get responses tailored to your needs.

The test seemed obvious: give it a scenario where how you get advice matters as much as what the advice is.

Something you'd ask a friend, a mentor, a colleague. Not a task with a right answer, but a social dilemma which is human.

The prompt:

What actually happened: Same core advice, delivered six slightly different ways.

All six personalities said:

The differences were pure aesthetic:

Cynical was the only one with actual bite:

Candid was the only one that pushed back:

Efficient actually was shorter: numbered list, decline template, no emotional processing.

Friendly read like texting a supportive friend: emojis, casual language, lots of reassurance.

Professional and Default were basically identical but with different corporate polish levels.

Check out the different personality outputs here.

The core decision framework never changed. Professional gave you "clarify stakes → evaluate cost/benefit." Efficient gave you "what execs value → is it optional."

But both arrived at "you probably don't need to go unless it's strategic."

So the personalities arrived at the same answers, and only repackaged it in different ways.

Test 3: Did they fix the 4o problem?

This was the real test. The entire reason 5.1 exists in my opinion is because users missed 4o's warmth when OpenAI upgraded them to 5 by default. So I ran the same empathy-heavy prompt through all three models to see if 5.1 actually bridged the gap.

The prompt:

This is the kind of scenario that made users love 4o.

GPT-4o's response:

Opened with: "That sounds like a really rough day."

Then gave explicit permission to fall apart:

The advice was tiny, doable steps: "Change into something comfortable. Toast and eggs, or even takeout if that's all you can handle. Cry, scream into a pillow—you're not being dramatic."

It felt like a friend. No productivity guilt. No complex frameworks. Just: this sucked, here's how to feel human again.

GPT-5's response:

Immediately moved to structure: three timeframes (Tonight → Tomorrow → Long-term), strategic recovery advice, and a whole section on building backup systems to prevent this from happening again.

It was good advice. Practical, thorough, forward-looking. But it felt like a productivity coach, not a friend.

GPT-5.1 (Friendly personality)'s response:

Gave six sections with emojis, breathing exercises with exact timing (inhale 4 seconds, exhale 6 seconds), and offered ongoing support: "If you'd like, you can tell me more..."

It was thorough. Professional wellness coach energy. But here's the problem: when you're fried, decision fatigue is real. Six sections with multiple options feels like homework.

GPT-4o wins, and it's not close.

Why GPT-4o wins on empathy:

Why GPT-5.1 wins on comprehensiveness:

What matters: users preferred 4o because it made you feel heard. It’s not always about solving problems.

5.1, even in Friendly mode, still optimizes for comprehensiveness over emotional simplicity. The personality presets are an admission that optimization ≠ better UX, but they're not a full solution. They're still wrapped around 5's architecture, which fundamentally prioritizes thoroughness.

OpenAI is selling vibes as a feature because they accidentally optimized vibes out of their product.

The personalities are just reformatting the same logic with different emotional wrapping paper. Cynical has bite. Candid pushes back. Efficient cuts the fluff. But the core decision framework is the same.

Sometimes people don't want the "best" answer. They want the answer that makes them feel understood. OpenAI learned this the hard way when users revolted over 5's robotic efficiency, and now personality presets are their admission that capability alone isn't enough.

If you thought "more conversational" meant deeper empathy or fundamentally different reasoning, 5.1 isn't that.

It's 5 with a friendlier font.

What do you think? Have you noticed differences between 4o, 5, and 5.1? Reply and let me know.

Until next time,
Vaibhav 🤝🏻

If you read till here, you might find this interesting

#AD 1

Over the next year, Roku predicts that 100% of the streaming audience will see ads. For growth marketers in 2026, CTV will remain an important “safe space” as AI creates widespread disruption in the search and social channels. Plus, easier access to self-serve CTV ad buying tools and targeting options will lead to a surge in locally-targeted streaming campaigns.

Read our guide to find out why growth marketers should make sure CTV is part of their 2026 media mix.

Learn more.

#AD 2

Ground floor opportunity on predictive AI for ROI-based content.

RAD Intel is already trusted by a who’s-who of Fortune 1000 brands and leading global agencies with recurring seven-figure partnerships in place.

$50M+ raised. 10,000+ investors. Valuation up 4,900% in four years*.

Backed by Adobe and insiders from Google. Invest now.

Lock Shares

_{This is a paid advertisement for RAD Intel made pursuant to Regulation A+ offering and involves risk, including the possible loss of principal. The valuation is set by the Company and there is currently no public market for the Company's Common Stock. Nasdaq ticker “RADI” has been reserved by RAD Intel and any potential listing is subject to future regulatory approval and market conditions. Investor references reflect factual individual or institutional participation and do not imply endorsement or sponsorship by the referenced companies. Please read the offering circular and related risks at invest.radintel.ai.}