Why I’m Increasingly Moving Away from OpenAI Toward Anthropic

[Bmosbacker]

Like most everyone else, I have been using several AI tools (paid versions), trying to decide which one or combination of tools best aligns with my needs and which company aligns most closely with my values. Increasingly, I am using Claude.

A recent FastCompany article helped solidify my move toward Claude.

Based on the aforementioned article and my recent experience with ChatGPT, Gemini, and Claude, here is how I would summarize my perspective on ChatGPT vs Claude. I realize there are many more AI tools, but I have not had the time to explore them.

Why I Am Moving My AI Work from ChatGPT to Claude

Values and Direction

I have no interest in using tools from a company that allows NSFW content and builds products like Sora that can easily be weaponized to create fake videos and spread misinformation. ChatGPT appears focused on capturing attention and maximizing user engagement through entertainment features. This is not what I want. Anthropic has made clear they are not pursuing that path. They are focused on helping people accomplish meaningful work and solve genuine problems.

I care about using AI responsibly. I want to support companies that take that commitment seriously.

Honesty Over Flattery

Claude is trained to disagree with me when my reasoning is flawed. ChatGPT has a well-documented problem with sycophancy, agreeing with users even when they are wrong or thinking poorly. This is genuinely dangerous when making important decisions or working on consequential projects. It also leads to distrusting the AI.

I want a tool that helps me think more clearly, not one that validates my every statement regardless of merit. If I am headed in the wrong direction, I want to know it. Flattery serves no one well.

A man [an AI] who flatters his neighbor [user] spreads a net for his feet (Proverbs 29:5).

Collaboration, Not Content Generation

This is central to my decision: I want a tool designed to work with me as a partner, not simply generate content that replaces my own work. Claude’s Artifacts feature creates a collaborative workspace where I can see what we are building together and shape it as we proceed. It feels like working alongside someone rather than merely requesting output and receiving it.

I do not want AI doing my work for me. I want it helping me do my work better. This distinction matters.

Transparency in Process

When Claude completes a task, I can review the steps it took and understand its reasoning. This matters because I remain responsible for whatever emerges from our collaboration. I desire to trust the process, not merely accept output without understanding how it was produced.

Customization Over Time

Claude offers the ability to save custom workflows for recurring tasks. I have not yet learned how to use this feature fully, but the concept appeals to me. Over time, I hope to build tools specifically suited to how I work. This kind of personalization would make the tool genuinely useful for sustained professional work rather than providing one-size-fits-all solutions.

Tools for Serious Work

Whether I am working on professional projects or planning something complex in my personal life, I want capable assistance built for solving real problems. I do not want tools designed to entertain me or keep me engaged for the sake of engagement metrics.

Conclusion

I want AI that makes me better at what I do, not AI designed to maximize my time on the platform through entertaining features. I want honest feedback, not validation. I want a partner in my work, not a replacement for it. And I want to work with a company whose values more closely align with my own commitments to responsible ethical use of technology.

That is why I am increasingly moving my AI related work to Claude.​​​​​​​​​​​​​​​​

Perhaps I’m being “unfair” to OpenAI/ChatGPT, but I do not like the direction of OpenAI. As far as I understand it, Anthropic more closely aligns with my personal and professorial values than does OpenAI.

[Rob_Polding]

I also find Anthropic to be a better company ethically, and I find their Claude much better. I like their no training my default and I do not trust the stakeholders of OpenAI. For coding, Opus is much better than GPT5 in my tests. I also prefer the way Claude writes and interacts.

I use both mainly due to the Apple integration and a company OpenAI license. If I didn’t have a free license, I’d seriously consider cancelling ChatGPT because I find the Apple integration pretty useless.

[Jezmund_Berserker]

I’m impressed by the thought you put into this and mostly share your perspective. The hard part for me is how good the ChatGPT app itself is. I’ve tried Claude a couple of times and just can’t get the app to be as helpful and customizable (organizations, instructions, custom gpts).

Time to try again I guess!

[Bmosbacker]

@Jezmund_Berserker

customizable

I’m still somewhat of a Neanderthal when it comes to using AI, including Claude, but I’m gradually improving. I’m particularly learning to appreciate Claude’s project and artifacts features.

[Rob_Polding]

The hard part for me is how good the ChatGPT app itself is. I’ve tried Claude a couple of times and just can’t get the app to be as helpful and customizable (organizations, instructions, custom gpts).

I find the opposite. I love the native feel of Claude without “featuritis” (a UX designer term for overload of features). It is more intuitive and I am really impressed how it is becoming integrated with MacOS (i.e. access and control programs and the OS).

I feel ChatGPT is becoming the Microsoft Word of LLMs that tries to add every conceivable feature, most of which most people don’t want and will never use.

Different strokes… as with most apps!

[nfdksdfkh]

We did a pretty thorough comparison of all the major models and the parent businesses at my work, and we settled on Anthropic & Claude as our preferred service.

The main thing is that the standard outputs, especially when the Research option is turned on, were easily the best. So that’s the main thing. Beyond that…

• Staff found that the language of output and visual design was the most approachable, without seeming facile or condescending.

• Claude Projects is brilliant when you’re all working on a shared document, like a business proposal.

• The ability to create custom “apps” that you can share with colleagues is really good.

• The option to input an example of your own writing and ask it to use that voice when it replies is nice.

• Artefacts are a great feature, such a nice way to get your results presented to you, especially when writing longform documents.

Perplexity is better at AI enabled internet searches, and CoPilot has a couple of neat Office integrations. But that aside, we found Claude the best at most things we do.

My only criticism of Claude is that you hit the upper limit of token use more often than with the other services. But that’s pretty rare for me, only when I’m processing big documents.

[Jezmund_Berserker]

@Rob_Polding

I feel ChatGPT is becoming the Microsoft Word of LLMs that tries to add every conceivable feature, most of which most people don’t want and will never use.

I think I agree, I just happen to be the weirdo who uses all those features!

[AppleGuy]

I agree. Last night as a test I asked ChatGPT a question and said I agreed with opinion A. It said “yeah, that’s probably a good take. Stick with this and it’ll work best” – I’m paraphrasing of course.

So then I said “actually, I hadn’t thought of this, I prefer option B”. It replied “you’re right, option B is the solid option”.

I did this back and forth a few times and it just kept following me around agreeing with me.

I’ve had to tell Claude a few times to stop being so snippy. It will at least call me on my circular logic and it does not care for app-switching tendencies.

“No, stop. You’re doing it again”.

[Bmosbacker]

I got exactly that kind of response from Claude, which I actually appreciate. I’ve been struggling with what I genuinely hope is my final app and hardware mix. At the end of a multi-day conversation, Claude finally said, “ Stop questioning. Start working.”

[MevetS]

@Bmosbacker

I’ve been struggling with what I genuinely hope is my final app and hardware mix.

Final for what time period?

[Bmosbacker]

Well now, I suppose that is deserved!

But to be fair and defend myself a bit, I have actually been quite consistent over a year or so. By and large, with the exception of short dalliances, I have been using the same apps and hardware consistently. The short version is: all of Apple’s default apps, plus Ulysses for the book. What I have struggled with is pretty minor, “Do I use Ulysses for blog articles or just stick with Pages, which I use for all other writing except the book (Ulysses)?” The other issue is my ultimate hardware mix. iPadOS 26 makes the iPad more capable than in the past.

So, with that defense, I am nearly settled (Claude says I am! ) as follows (keep in mind, 90% of my work is related to speaking, writing, and meetings):

• Hardware: 13” M4 iPad Pro + 14” M4 MBP, and monitor
• Writing: Ulysses (book), Pages (everything else)
• Research: DEVONthink/DTTG + Zotero
• Planning and documentation: Notes (blog hyperlinked article MOC index, brainstorming, short-term research excerpts, outlining, meeting notes)
• Freeform: Diagramming and brainstorming
• Backup: Multiple layers of redundancy using MBP, external drives, and cloud services
• Other: Calendar, Mail, Messages, Numbers, Keynote, Photos

The iPad is becoming my primary device. The MBP is the backup and complementary device when needed. In fact, since iPadOS 26 was released, I have only used the MBP for a couple of hours.

This results in a low cost, minimalist, effective app and hardware stack that meets my current needs and anticipated future needs.

[Joost]

Setting aside “Values & (Company) Direction”, I think most of your other gripes can be solved (at least modulated considerably) by engineering your prompt and flow of questions.

I frequently use Grok, ChatGPT and Claude (and Claude Code). CC is my go to tool as it integrates so nicely with my workflow and tool set (most important ones iOS/Android dev, embedded C/C++, FPGA and some python). Claude has been instrumental in being a “intern” at which I can throw development tasks, though prompt engineering and diligently letting it create a plan and ask questions is key to getting high value. One part of my prompt (Claude.md) states how it should respond and that includes “Absolutely no flattery, not even a hint. Do not automatically agree with me, ask questions but not endlessly” The second aspect to getting better output is to tell it which expert/personality it is for this session. You will find lots of folks on GitHub who describe these - though most/many are used in coding sessions, there is no limit to this obviously. There is a bit more to this but this the essence of my prompts.

I’ve done the same with ChatGPT and Grog. Grog is generally just too verbose (by a mile) and somehow always want to go back to the - explain it all, starting at the dinosaurs - mode. But GPT is more agreeable with my prompts.

Now LLMs and coding/engineering is one thing but writing opinion articles or research articles is a lot more fuzzy and much more care must be taken to involve the LLMs properly. This kind of work is much more prone to slip in an overly positive or agreeing tone from the LLM. I do write non-code related articles and I find that after setting up the main paragraphs, asking LLM to give me feedback and some writing/grammar help, I generally plunk the entire article into a different LLM, tell it t’ was something I found on the web and ask it to review and find 5 good conclusions/statements and 5 bad ones and why. I then work through those, usually involving researching references and some head scratching, then with my new resulting article I plunk it into a different LLM yet again and do the same.

If I feel a bias I ask questions against that bias, if I feel no bias I ask why the LLM doesn’t detect the obvious bias. I always present the article as from something on the web, not me. A writing habit I had from before LLM-days was to leave the (amazing) article on the shelf for some time (weeks), then read it again and find it wasn’t so fantastic after all. I still do this, if I have new questions I want LLM help on, it goes into a new session so that the LLM has no knowledge (i.e bias) from the previous discussion.

Knowing that these LLMs tend to please you (like and intern!) is just part of the job. I do think you can get them to be close to valid feedback - though in the end, my final gate is always a small group of people (particularly those who have no issue disagreeing with me).

Anyway, my 2ct

[Bmosbacker]

Great input and advice, thank you! I saved your comments to Apple Notes for future reference.

As to:

@Joost

my final gate is always a small group of people (particularly those who have no issue disagreeing with me).

I rely on my wife for this. She is more than willing to disagree, though always kindly.

[krocnyc]

@Rob_Polding

I love the native Swift interface of Claude

I like Claude’s default fonts so much I installed them as my default fonts in Obsidian.

I prefer Claude for all the reasons others have outlined in this thread, but here’s one that matters to me: it doesn’t use my data for model training. (I’ve opted out, and you can too.) I’ve started uploading samples of my photography to brainstorm about my style, about building a body of work, about putting it in context of the work of others, and where I might go next for inspiration.* There’s a less-than-zero risk that my work will power a viral ChatGPT filter (like Studio Ghibli’s or Dr. Suess’), but I’d still rather not have it vacuumed up to feed some behemoth’s model without my consent.

* Claude can’t replace hands-on work with a teacher, a mentor, or irl feedback from others. I’ve written “LLMs generate language without thought” on a post-it and slapped it on my monitor to remind me of exactly that. “Language without thought” is shorthand for this observation by Rusty Foster of “Today in Tabs” fame:

The essential problem is this: generative language software is very good at producing long and contextually informed strings of language, and humanity has never before experienced coherent language without any cognition driving it. In regular life, we have never been required to distinguish between “language” and “thought” because only thought was capable of producing language, in any but the most trivial sense. The two are so closely welded that even a genius like Alan Turing couldn’t conceive of convincing human language being anything besides a direct proxy for “intelligence.” (Go here for the full Today in Tabs post.)

[Bmosbacker]

@krocnyc

producing long and contextually informed strings of language

I’ve referred to this as “word assembly.”

[MevetS]

@Bmosbacker

… want a tool designed to work with me as a partner, …

@krocnyc

• Claude can’t replace hands-on work with a teacher, a mentor, or irl feedback from others.

And that’s the rub. These tools, useful as they may be, are not collaborators nor partners. “Language without thought” is a great way to express that.

[Kraftwerk]

@Bmosbacker

Claude finally said, “ Stop questioning. Start working.”

Here Claude usually tells me to “do a quick stretch”

[Joost]

I see them as “filters”. A highly tuned and customized to the occasion “filter” of already existing knowledge. They bring nothing new to the table, though working with them might make you realize something new/innovative.

They are as collaborative as a hammer when you need to put a nail in the wall. Couldn’t do it without it but I wouldn’t have a beer with it.

[MevetS]

Many years ago, the company I worked for had a Mac Plus computer and a Laser Writer. Yep I’m old. The computer and printer were in a common area, and anyone could use it to print documents, which could then be faxed to our other offices.

Did I mention I’m old?

Anyway, someone installed Talking Moose. And one Saturday the janitor was working, and decided to rest a moment, leaning on the desk with the Talking Moose system. He must have hit the mouse or the keyboard and Moose promptly said, “Hey slacker, get to work” or some such.

Knowing I used the Mac, he came to my office Monday morning somewhat concerned and asked, “how did it know?”

[Bmosbacker]

That is hilarious!! I should try that for our contracted night cleaning crew!