GPT-5: The Good, The Bad, and The Tasteless

Good morning!

Welcome to The Daily Grind for Friday, August 8.

We’re coming in a little late today, but I’m excited to share The Daily Grind’s brand new website and newsletter on Beehiiv!

You can also access all our articles via our fancy new URL: dailygrind.news. Tell all your friends.

Today’s headline is about the release of GPT-5: The Good, the Bad, and the Tasteless.

Then our One Page, we’ll read from one of the literary world’s great editors to understand how taste is developed.

Finally, our One Question will ask you to evaluate your own taste: where you trust it and where you don’t.

Let’s get started!

📰 One Headline: OpenAI releases GPT-5

Days after OpenAI, Claude, Google, and Alibaba all made relatively small model releases, OpenAI followed up with a massive splash. GPT-5 is now available to all users—paid and free—for use in ChatGPT and in the OpenAI API.

The team at Every has been testing GPT-5 for weeks and the results are surprisingly mixed. Here’s a breakdown:

The Good: Fast, Simple, Useful, Powerful… and Cheap

• ChatGPT Simplified: In what Dan Shipper calls an “Apple-esque” move, OpenAI cleaned up the ChatGPT interface by hiding the thinking modes like study, deep research, and agent mode. ChatGPT will now decide on its own which mode is most appropriate based on the query or task. The new ChatGPT interface is very clean… almost like a legendary Apple designer had a hand in it.

• Fast and cheap: For basic queries, GPT-5 is lightning fast. For developers and builders, OpenAI also just reset the market on price. GPT-5-mini undercut Gemini 2.5 Flash, Google’s most cost effective model, at $0.25 per million tokens vs. $0.30. For its main model, GPT-5 comes in at 12x cheaper than Claude 4 Opus.

• Optimized for health queries: Justine Moore, a partner at a16z, pointed out on X that GPT-5 is more reliable for health-related questions. Anyone who has felt the frustration and occasional dread of searching symptoms on webMD should be very pleased.

• Writing and collaborative coding: The Every team describes GPT-5’s writing voice as “nuanced and expressive,” and less likely to use common AI idioms while writing. On the coding side, GPT-5 is great at helping developers troubleshoot and debug, helping one Every engineer solve a bug that had vexed him for 4 days.

• Deep Research: OpenAI put a lot of effort to improve GPT-5 reasoning abilities. It thinks more deeply and comprehensively than any previous model. And it’s responsive to your requests to “think deeper” on a subject.

The Bad: Autonomy, Confidence, and Taste

• Autonomous coding and agentic tasks: Where GPT-5 excels with collaborative coding, the Every team found that it struggled to work on its own. “It's more cautious than Opus 4.1 and isn't as comfortable working independently for long periods in our testing,” wrote Every CEO Dan Shipper.

• Personality and style: When the Every team asked to one-shot an interactive video game and a music product app, GPT-5 was strong on technical details but light on aesthetics.

• Evaluating good writing: GPT-5 could write well enough on its own, but when asked to evaluate another person’s writing, Every editors found it was inconsistent and occasionally just wrong.

In summary, GPT-5 is a braniac genius that’s still trying to find its voice. It impressive with sheer intellectual horsepower, but it lacks taste.

Dan Shipper summarized the state of play this morning on X:

Read Every's full evaluation, it’s worth your time, and then start using GPT-5 for yourself.

🔗 A Few Good Links

GPT-5’s release sucked up most of the oxygen yesterday, but here are a few more stories to explore:

• Bloomberg: SoftBank purchased Foxconn's EV plant in Ohio for $375M, a move aimed at kickstarting its $500B Stargate data center project with OpenAI and Oracle

• Wall Street Journal: An analysis of 96K public ChatGPT transcripts finds 100+ lengthy chats, of which dozens exhibited delusional traits where ChatGPT reinforced users' false claims

• TechCrunch: Windsurf's gross margins are “very negative”; many believe the same margin pressure is impacting Cursor, Lovable, Replit, and other vibe coding tools

📚 One Page: Avid Reader by Robert Gottlieb

When Merriam-Webster announces its Word of the Year, I won’t be surprised if it’s “taste.”

Taste is the human antidote to AI-generated slop, and according to some, you either have it or you don’t (and apparently, GPT-5 does not.)

I don’t think that’s true. Taste is developed through a process of thoughtful consumption, evaluation, and, importantly, creation.

You’ve probably never heard of Robert Gottlieb, but you certainly know some of the books he edited: Catch-22, Power Broker, Andromeda Strain, and several works of Toni Morrison.

This passage, from Gottlieb’s autobiography Avid Reader, recounts the time Gottlieb spent working with Joseph Heller on Catch-22. It encapsulates the magic and challenge of good editing. Three things stand out to me:

1. The very loose, unstructured way good writing is crafted

2. Robert’s own confession of not knowing where his taste came from (hint: it’s in the title of his book)

3. How good editing disappears from sight, which is why it’s so hard to grasp for new or non-editors.

Avid Reader is one of my favorite books on writing.

❓ One Question: Where do you have “Taste?”

As Gottlieb shared, we all have “taste” in some arenas of life, but not others.

In what arenas or areas of work do you trust your “taste?”

Trusting your taste in something is probably a signal that you have a strong interest and competence in a subject. Follow that taste.

Like Gottlieb, I trust trust my judgement as a reader and editor. I know what I like and what I don’t, and I take time to understand why.

For me, taste starts as a feeling—it’s either delightful or grating. Then I investigate further. Why is this thing delightful? Why is it grating? I search my mind for analogies or examples until I understand the feeling. Then I articulate it. Then I put the lesson to work in my own writing.

Knowing where you trust your taste can also help you identify the areas where you need to find trusted tastemakers. When it comes to AI models, I trust Every’s taste more than my own. They are in new models every day, doing the work, taste-testing every development.

So where do you trust your taste?

🗳️ Wrap Up and Feedback:

That’s it for today’s Daily Grind!

Today is the first day publishing The Daily Grind on Beehiiv. What did you think? How was your experience?