GPT-4.5 Wants to Be Your Friend

Hello, and happy Sunday! This week Dan Shipper and Alex Duffy got an early look at OpenAI’s newest model, ChatGPT-4.5. Their assessment: It’s friendlier, slower, pricier, and surprisingly human. Read on for our complete review, as well as everything else we published.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

Release notes 

GPT-4.5 arrives

This week, OpenAI dropped GPT-4.5 as a research preview for ChatGPT Pro users, and our first impressions are…mixed. We published our complete thoughts several hours after it went live.

No, it's not the jaw-dropping leap in intelligence we’d hoped for, especially given a reported order-of-magnitude increase in the cost of training the model compared to previous versions. Some benchmarks improved and hallucinations decreased slightly, but it's no slam dunk. Instead, OpenAI leaned heavily into emotional intelligence, creativity, and conversational warmth. 

If GPT-4o feels like an efficient assistant, GPT-4.5 is trying to be your empathetic, opinionated bestie.

The good: It writes better. Responses have rhythm and clarity, and feel surprisingly human—OpenAI calls it "Orion prose," optimized for spoken conversation. This could be huge for voice-based AI interactions, especially if you consider Sesame’s latest voice AI demo (seriously, check this out—Evan Armstrong called it like having “a glitchy phone call with a really flirty person”). 

Emotionally, 4.5 is a noticeable step up, capable of genuinely funny, comforting, and insightful interactions. Sam Altman tweeted that talking with GPT-4.5 feels "like talking to a thoughtful person," and frankly, he’s not wrong.

GPT-4.5 even “yo mama”’d me:

Source: ChatGPT-4.5. Courtesy of Alex Duffy.

The bad: It’s SLOW. Painfully slow. And restricted to the $200 a month ChatGPT Pro tier. Altman admitted on X that they're facing a severe GPU shortage due to surging growth. The company will  soon roll out tens of thousands of additional GPUs, but for now, using GPT-4.5 can test your patience.

And speaking of GPU shortages, get ready for sticker shock. The GPT-4.5 API cost is approximately 15-20 times more than GPT-4o and significantly more than competitors—many of which outscore GPT-4.5 in multiple benchmarks. 

One more caveat: GPT-4.5 struggles with instructions. It's opinionated, so it also deviates from directly answering your prompts in an attempt to come up with creative responses. This might initially frustrate, but given time, we found ourselves warming to its quirks. 

This mixed bag has led to some disappointment from the tech community, many of whom have been eagerly awaiting a next-generation model since GPT-4's release almost two years ago. Despite these bumps, GPT-4.5 is fascinating precisely because it's a different kind of intelligence. It won't dominate the benchmarks, but it may become your favorite conversational companion—especially once paired with advanced voice interfaces. 

Why OpenAI chose to move in this direction starts to make sense when you zoom out a little: Big model providers are starting to pick their lanes. Anthropic and Claude are hyper-optimizing for code to woo engineers, while OpenAI is hoping GPT can become the go-to AI assistant for everyone. 

For a deeper dive into what makes GPT-4.5 tick—including surprising personality test results, detailed hallucination checks, and hands-on writing tests—check out our full review.—Alex Duffy

Now, next, nixed

The current state of LLM evaluation benchmarks.

Knowledge base

🔏 "AI Turned Me Into a Content Agency of One" by Katie Parrott/Working Overtime: Katie Parrott has become a one-woman content factory with the help of AI, taking on workloads that would've kept four writers busy for months. Her six-part AI workflow is a huge force multiplier—but it also raises questions about the ethics of AI-enabled efficiency. Read this for a reflective guide on bringing AI closer to your work. 

🔏 "I Created a Hacker News Simulator to Reverse-engineer Virality" by Michael Taylor/Also True for Humans: Michael Taylor scraped the histories of 81 real Hacker News commentators to build his own AI focus group—and found that getting AI to be mean might get you better insights. (Did you know ChatGPT could swear?) Read this if you're ready to revamp your customer research.

🎧 "He Built an AI Audience Simulator. It’s the Future of Customer Research." by Dan Shipper/AI & I: The future of customer research is here, and it doesn't involve a single awkward focus group. Dan sits down with prompt engineering wizard Michael, whose tool Rally creates AI personas that mimic real audiences with uncanny accuracy. Listen to this if you want to learn how to make AI into a better judge of your work. Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

"GPT-4.5 Won’t Blow Your Mind. It Might Befriend It Instead." by Dan Shipper and Alex Duffy/Chain of Thought: OpenAI's latest model isn't a revolutionary leap we expected, but after extensive testing, Dan and Alex found that GPT-4.5 trades some instruction-following precision for more human-like responses—it's more opinionated, emotionally intelligent, and writes with a natural flow that previous models lacked. Read this for an early readout on OpenAI’s latest. 

“Playing ‘Civilization 7’ Is Like Being a Ruthless Corporate Consultant.” by Evan Armstrong/Napkin Math: Mark Zuckerberg logged over 1,000 hours in the first six games of Civilization. Now, the newly seventh version is out, and in it, any leader can conquer any land—not unlike today’s tech industry. Read this for Evan’s analysis on the strategies that gamers and CEOs share.—Aleena Vigoda

From Every Studio

Get early access to multiple accounts on Cora: Since we launched our inbox-human email experience, users have been asking to connect more than one account. Starting next week, you’ll be able to add multiple email addresses to your Cora mailbox. If you’re reading this and eager to link up your accounts, send a DM to Cora’s general manager, Kieran Klaassen, to get set up early. 

Link-in-Spirals: Here’s an update that makes it easier to get the most out of Spiral, our tool that remixes content, like podcasts and articles, into various formats, like social posts. You can now paste any article link directly into your Spiral input box—no more copying and pasting text manually. Let us know what you think, request a feature, or keep up to date with our changelog here. 

What’s next for Sparkle: Sparkle cleans your desktop, documents, and downloads folders to keep your digital workspace tidy. Coming up for Sparkle v2: giving users more say in how Sparkle organizes your files—including a slick dashboard for file stats, even better external drive features, and more intuitive UI changes.

GeoGuessr for designers: Test your color chops with Color Guesser, made by Every creative lead Lucas Crespo. GeoGuessr lands you on a random place in the world and asks you to find where you are on the map. Color Guesser challenges you to pinpoint the right hex code on a color wheel—the closer you get, the more points. Try out the MVP and let Lucas know your thoughts.—AV

Alignment 

Digital slavery. Byung-Chul Han's The Burnout Society is a damning indictment of modern day living. In fewer than 75 pages Han was able to put a name to what I felt but couldn't describe—this constant drive toward overactivity. He argues we have become hyperactive, perpetually scanning our environments like animals in the wilderness because we are always on. We’re hyper-alert to emails, notifications, news cycles, and productivity metrics, stuck in a self-imposed state of vigilance that our bodies and minds were never designed to sustain. Indeed, the premise of the book is that we now fundamentally work ourselves harder than any external taskmaster would dare. I would argue that this same frantic energy powers our AI obsession. Each week brings another model to master, another system demanding attention. We build machines to accelerate cognition while neglecting our own capacity to reflect and contemplate. My solution is quite simple. I block off time to do absolutely nothing with the same determination I protect my most critical meetings. It's not complex, but I increasingly feel that choosing stillness feels like rebellion.—Ashwin Sharma

---

Hallucination

An exploration of cinder block speakers.

That’s all for this week! Be sure to follow Every on X at @every and on LinkedIn. 

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.