OpenAI Just Dropped GPT-5 (And I'm Still Figuring Out What to Think)

Okay, let's talk about OpenAI's GPT-5 release. I'll be honest—from my initial poking around, it hasn't exactly blown me away yet, but there's clearly a lot more here than meets the eye. Sam Altman's making big claims, early testers are reporting some wild stuff, and I'm somewhere in that awkward middle ground of "this seems important but I need more time to figure out why."

Sam Altman dropped the news Wednesday, calling it "a significant step along the path to AGI." Now, before you roll your eyes at another "we're almost at AGI" announcement (trust me, I've heard them all), this one actually feels different. Not because we've reached some sci-fi singularity, but because of how it fundamentally changes the way we interact with AI.

The "Just Handle It" Approach

Here's what's supposed to be different: GPT-5 doesn't make you choose between models anymore. Remember the old days (aka last month) when you had to decide between "fast but dumb" and "slow but smart" options? Those confusingly named model variants that made you feel like you needed a computer science degree just to pick the right one?

That's gone. GPT-5 claims to automatically route your request to whatever computational power it thinks you need.

Ethan Mollick (one of the biggest AI thinkers in the space) has been putting this through its paces, and his findings are way more interesting than my basic experiments. According to his testing, GPT-5 didn't just build what he requested—it started suggesting improvements, adding features he didn't ask for, and basically acting like an overachiever intern who can't help but go above and beyond.

The Arbitrariness Problem (Or: Why I'm Slightly Concerned)

But here's where Mollick's testing gets interesting—and where my initial skepticism might be warranted.

According to Mollick, GPT-5's model selection seems somewhat random. In his SVG creation tests, identical requests would sometimes get basic treatment (meh results) and sometimes trigger the full-power reasoning mode (amazing results). The difference? Completely unpredictable.

Apparently, adding phrases like "think hard" to your prompts increases your chances of getting the smart model. Which feels like having to say "please" to a vending machine to get it to actually dispense your snack. It works, but it's weird that it has to.

I haven't run enough tests to confirm this myself, but if true, this inconsistency could be genuinely problematic for business use.

The Numbers Game

Let's talk specs for a moment (because someone has to). GPT-5 comes with a 256,000-token context window—up from 200,000—which means it can handle longer conversations and bigger documents without forgetting what you were talking about three prompts ago.

On coding benchmarks, it scored 74.9% on SWE-Bench Verified and 88% on Aider Polyglot. For non-technical folks, that basically means it got really good at understanding and writing code across multiple programming languages.

The hallucination rates dropped by 26% compared to GPT-4o, and for the reasoning variant, by 65% compared to the previous model. Translation: it's less likely to confidently tell you complete nonsense, which is always a win in my book.

The Pricing Reality Check

OpenAI rolled out a three-tier structure that's clearly designed to funnel people toward paid plans:

• Free tier: Access to GPT-5 with basic limits (aka "here's a taste")

• Plus ($20/month): Same models, higher limits (the sweet spot for most of us)

• Pro ($200/month): Unlimited access to all variants including the fancy reasoning models

For API users, it starts at $1.25 per million input tokens, with a lightweight version at $0.05 per million tokens—which actually undercuts Google's pricing. Rare to see OpenAI being the cheaper option these days.

The Overeager Assistant Problem

Here's what's supposedly different about GPT-5: it doesn't wait to be asked. When Mollick requested startup ideas, the system didn't just give him concepts—it autonomously created landing pages, financial projections, and marketing copy. Nobody asked for those, but GPT-5 was like "while I'm at it, here's everything else you probably need."

In his coding applications, it continuously added features without being told to—dynamic lighting, camera controls, save systems. It's like having a developer who can't help but keep polishing until they've built something way more complex than you originally wanted.

Whether this is incredibly helpful or mildly annoying probably depends on what you're trying to accomplish (and whether you actually have time for the AI to go down rabbit holes).

What This Might Mean for Our Businesses (Once We Figure It Out)

As someone who's still in the early stages of understanding what GPT-5 actually brings to the table, I'm approaching the business implications with appropriate caution.

The Potential Upside: If the proactive behavior and automatic optimization actually work as advertised, this could lower the barrier to entry for team members who aren't AI experts. Less time spent figuring out which model to use, more time getting work done.

The Reality Check: Based on my initial testing, I'm not seeing dramatic improvements over GPT-4o or o3 for basic tasks. Maybe I need to push harder, or maybe the real benefits only show up in complex, multi-step scenarios that I haven't explored yet.

The Control Question: The bigger concern is whether giving up granular control over AI behavior actually helps or hurts productivity. In my limited experience so far, I can't tell if the automatic routing is making better decisions than I would—which is either a sign that it's working seamlessly, or that the differences aren't as meaningful as the marketing suggests.

The Bigger Picture Shift

With nearly 700 million weekly users, OpenAI clearly feels confident enough to prioritize automated optimization over user control. We're moving from "tell the AI exactly what to do" to "give the AI a general direction and let it figure out the details."

This represents a fundamental shift in the human-AI relationship. We're becoming more like managers giving high-level guidance rather than technicians providing detailed instructions.

For those of us who've gotten comfortable with the old way of doing things, this transition might feel uncomfortable. The question isn't whether GPT-5 is more capable—early signs suggest it definitely is—but whether giving up granular control aligns with how we want to run our businesses.

• Walmart, Uber, and Instacart are cashing in on high-margin ad dollars as retail media becomes a $10B+ business reshaping the future of commerce and brand strategy

• A growing number of teens are turning to AI for support.

• Social media surges, search behavior shifts, and online video dominates TV as the latest Digital 2025 report drops with 540+ pages of insights to decode the evolving digital pulse

• An AI model used in a Vogue x Guess ad sparks outrage

  

• A new era of AI and Facebook/Meta! What do people have to say to this about privacy, control… manipulation?

  

• 2025 State of AI in Retail

• Pharma brands are embracing AI influencers 

• ChatGPT is now pulling in $1B per month with 700M+ weekly users as OpenAI hits a $12B run rate and ramps up spending with $8B projected in 2025

• Backed by Amazon, Fable’s Showrunner lets users generate their own sitcoms with simple prompts turning viewers into creators inside animated, episodic story worlds

• The T-shaped marketer is out and the AI-shaped marketer is in

🎨 OpenArt - App store for AI image apps.

🖼️ Moodboard Creator - AI-driven branding kickstarter

⚙️ Trellis: AI-powered eCommerce merchandising software, optimising Amazon and Walmart ads for product content, pricing, placement, and promotion.

📊 Simple Analytics: Get insights about your business by chatting with your analytics in real-time using AI.

🌐 Airo by GoDaddy - Get a website that literally writes itself. Logo, website, email and social campaigns, you name it.

❄️ Snowball - Simplify your content creation with AI assistance

😍It was great, all I needed

🙂Good, not great

🙄It sucked