Introducing GPT-5: Smarter, sharper, and more sarcastic

By now, most marketers know the cycle. Big AI model launches. Benchmark bragging. Glossy demos. A promise that this time, your prompts will be met with flawless prose and zero hallucinations. Then, within days, you’re knee-deep in inexplicable weirdness and wondering if the upgrade was worth it.

OpenAI’s new GPT-5 is here, and it’s aiming to break that cycle — or at least smooth some of the rougher edges. I want to say one thing up front – this is not an epochal, universe shifting moment. It appears to be a substantive upgrade, but more incremental than exponential, at least based on my testing so far. If this observation bears out, that has interesting implications relating to the speed of AI progress. 

It’s definitely a more unified experience than the old constellation of GPT models, which often felt like switching between three different interns mid-project. Now, GPT-5 and GPT-5 Thinking behave more like different gears on the same bike. You still need to pedal, but the ride is smoother. And in the spirit of putting it through its paces — and hitting deadline — I’ve used GPT5 to help write this article.

First impressions: Less chaos, more control

One of GPT-5’s most obvious wins is coherence. No more realising halfway through a coffee shop search that you’re still in “o3 mode” and need to restart the query. The structure now mirrors Google’s Gemini 2.5 approach — a core model and a thinking variant — and marketers used to juggling tools will appreciate not having to keep a mental map of which version they’re talking to.

Benchmark-wise, OpenAI claims the expected “better at maths, coding, science” upgrades, plus improved writing chops — “compelling, resonant writing with literary depth and rhythm,” in their words. Whether it’s actually poetry-ready is debatable, but the underlying tone and rhythm are less stilted.

Then there’s the new “Customise GPT” setting — a quiet revolution for anyone tired of the default chirpy sycophancy. You can pick personalities:

• Default: Cheerful and adaptive (read: “Sure, boss!” to everything).
• Cynic: Critical and sarcastic.
• Robot: Efficient and blunt.
• Listener: Thoughtful and supportive.
• Nerd: Exploratory and enthusiastic.

After brief experiments (Cynic was too flippant for client work; Nerd felt like chatting to a hyper caffeinated grad student), I settled on Listener, then bolted on prompts to make it more of a straight shooter.

(Click to enlarge)

The big claim: Fewer hallucinations

If there’s a ghost that haunts every generative AI model, it’s hallucination — outputs that sound authoritative but are pure invention. GPT-5’s headline boast is a 45% drop in factual errors compared to GPT-4o, and an 80% drop in Thinking mode versus the much-maligned o3.

Two years ago, I tested GPT by asking for academic papers on journalism and misinformation for my thesis. It obligingly made them up, attributing them to real authors. GPT-5, under the same test, served up an actual (and relevant) reading list. A few citations routed me via Wikipedia instead of the original source, but in those cases the real paper was linked in the footnotes — a clear improvement.

Is it really 40-plus percent better? Hard to say from one morning’s work, but the early signs are promising. That said, as you’ll see later, there’s evidence the hallucination problem isn’t entirely solved.

A practical test: Marketing my new business

As a benchmark, I gave GPT-5 a job that matters: generating marketing ideas for my newly launched consultancy, The AI Training Company. (Side note: this article now doubles a soft launch for my business and its new website. I haven’t spoken publicly until now.) 

The brief:

You are a creative director specialising in helping small businesses. Give me five original ideas for marketing my new company, the AI Training Company. My goal is to reach a broad range of clients who need help training their executives, boards and companies in how to use AI.

I ran the same brief through Gemini 2.5 Pro for comparison.

Results? Mixed. GPT-5’s Thinking and Pro variants edged out Gemini for originality. Both converged on a neat concept — an “AI Fire Drill” for boards, simulating high-stakes AI incidents to teach practical safety principles.

Here’s the side-by-side table:

The full grid made it clear: GPT-5 was better at riffing on distinctive concepts, though not always the fastest to package them neatly.

Idea generation is one thing; bringing it to life is another. I asked GPT-5 Thinking to turn the “10-Minute Board Brief” into an interactive web module. Gemini nailed it in one go — clean design, plain-English copy, balanced layout. 

GPT-5 Thinking needed several nudges on fonts, colours, and logo placement, and the final draft skewed a little too technical. This wasn’t a deal-breaker, but in a head-to-head for polished, ready-to-ship collateral, Gemini still wins.

Writing this article: How the models performed

To write this piece, I gave my notes to GPT-5, GPT-5 Thinking, and Gemini 2.5 Pro. The prompt (plus some extra tone and style notes) was:

Your job is to write an article in my style for the marketing website Mumbrella. It is about the introduction of GPT-5. For the article contents, draw from the Docs file “GPT5 Notes”. Also supplement it with information from OpenAI’s announcement (link). Use the file “ChatGPT Agent Article” as a reference for my style of writing.

GPT-5 Thinking was not the right choice. Rather than supplementing the notes with information from OpenAI’s announcement, it created whole new sections. It also went heavy on technical notes (“For developers and martech teams, the API gained new dials: a verbosity parameter…”), which was not what I wanted. 

GPT-5 itself did a really good job, and that’s what you’re mostly seeing here. The copy is lively and closely reflected the tone and substance of my notes. So at first blush, it seemes that GPT-5 vanilla might be a better choice for writing tasks. 

But one big problem — it hallucinated a section saying I had created a Custom GPT to write the article, when I had done no such thing.

Gemini 2.5 Pro was also excellent. There was little to separate it and GPT-5. I felt a slight preference for Gemini’s version but can’t quite put my finger on why — maybe just familiarity. I’m calling it a draw.

For marketers, the bottom line

For everyday content tasks, GPT-5’s improvements in accuracy, personality control, and coherence will make it easier to trust — and faster to brief. The reduced hallucination rate could mean fewer frantic corrections in the middle of a campaign. The personality tuning lets you shape a model that suits your brand voice rather than fighting against default cheeriness. There’s also a claim of enhanced memory, which I have not been able to test yet.

It’s not flawless and I also don’t think it’s breathtaking leap forward, at least not in the tasks I undertook. Gemini seemed to match it most of the time, and the “literary depth” claim should be taken with a pinch of salt. If the hallucination rate claims are true, that may be the most meaningful contribution of this model. 

Overall I’d say this is a major step up in user experience for GPT users. It performed well on every task that I gave it and OpenAI’s confusing model salad problem has been solved. I’ll be hammering GPT5 for the next week and I’ll come back with any further insights beyond the two-hour spin from this morning. 

Shaun Davies is the principal and founder of The AI Training Company and a former Microsoft product manager.

Got a tip?

topics ChatGPT ChatGPT Agent OpenAI Sam Altman

Share Tweet Share

Have your say

Login with:

Continue with LinkedIn

or

Comment anonymously

Your comment will be marked as unverified

Name

Comment

Δ