Ask HN: If you've used GPT-4-Turbo and Claude Opus, which do you prefer? | Hacker News Reader

Ask HN: If you've used GPT-4-Turbo and Claude Opus, which do you prefer?

113 pointstikkun101 comments

Which do you prefer and what do you prefer about it?

GPT-4-Turbo is the default model in ChatGPT Plus.

Claude Opus. GPT-4 gives sensible answers to basic questions, but takes serious persuading to produce useful output on non-trivial work. Opus can be used as an integral part of an engineering workflow and only takes 2-3 tries to get from ill-formed query to working product.

This works particularly well when you copy the relevant excerpt from a project, dump it in, and say "Change X to Y, showing only the key modifications and where to put them". Typically it understands the aim and accomplishes the task in the way you intended, and it knows how to be concise yet precise.

Is there a noticeable difference between Opus and the free Sonnet?

Yes

Even then, Sonnet (free) is much better than GPT-3.5 (also free)

I use Sonnet unless I need something actually complex

Claude 3 Opus by a wide margin. I'm a regular GPT-4 user who had tried Claude 2, and went into Claude 3 with muted expectations. I was shocked with how much more capable Claude 3 Opus was compared to GPT-4, it's not even close for my work (optimization, programming/algorithms). I asked 20 questions and Claude 3 solved all of them where GPT-4 failed on all of them. What's more surprising to me is that I don't remember GPT-4 being this bad, I was similarly impressed when GPT-4 was released. The disparity was significant enough for me to reconsider my subscription to OpenAI, but the Browsing capability as well as the Android app kept me. I use Opus for my work now though, by default, and fall back on GPT.

GPT4 has been significantly neutered. Personally I suspect it's a combination of model updates that aren't that good, but also resources being moved over to GPT5. I suspect there's a culture of "no roll-back".

If you go to the OpenAI playground and try out the original model, gpt-4-32k-0314, you'll see a dramatic difference in responses, especially for coding.

The turbo model is particularly bad.

I fear that Claude will become neutered, it's already refusing a prompt that I was using a few days ago without issue, infact I can still go back to the old chat and continue it, but restarting it Claude refuses.

For coding? Claude Opus.

For me, GPT-4-Turbo is significantly worse than even GPT-3.5: the former is much better at providing context for its answers (even erring on the too-verbose side), but then comes up with a pointless solution that it can't be dissuaded to change, even if its predecessor gets it right-ish.

Compared to both these GPT versions, Claude 3 (even though I have to use a proxy to pretend I'm in Nigeria...) is much more 'to the point' and seems more 'willing' to amend answers that don't go in the right direction, as opposed to simply backtracking and proposing a completely new solution.

But having to pare down the context of a question significantly remains a huge issue for all models, and I think this is their Achilles heel. Until you can feed a model your entire project, including any dependencies, and it can answer any questions in the that full context, the work required to retrofit useful answers is just too much to justify the expense.

> I have to use a proxy to pretend I'm in Nigeria

Could you explain more?

+1 I am genuinely confused

I'm located in the EU, where Claude is unavailable: https://www.anthropic.com/claude-ai-locations.

So, since I have access to some IPs in Nigeria, I used those to (brazenly, possibly illegally!) evaluate their services (and no, the recent sea cable cuts don't help, but don't seem to affect my African upstreams too badly).

> you can feed a model your entire project, including any dependencies, and it can answer any questions in the that full context,

I've tried Gemini 1.5 with ~1M tokens and it took >90sec to answer anything.

How did you get access to Gemini 1.5, I thought it wasn't available for general access yet?

There's a waiting list for people who want to try it for free.

Ah, should have probably googled before asking. Thanks!

I prefer Claude, but for a non-performance reason.

ChatGPT has a writing style that is recognizable. So, Claude outputs don’t seem as AI generated, but probably only because ChatGPT is more popular.

Not available in Eastern Europe.

it's on poe

what is poe?

(I'm sure you don't mean power-over-ethernet)

They mean poe.com, the LLM aggregator. The LMsys arena also has worldwide Claude access at the moment, that's on chat.lmsys.org, direct chat.

Yes, thanks for clarification.

You can probably use the API, I always use the API even with GPT-4 as it is cheaper for the vast majority of users and probably faster.

It’s basically not available in Europe at all. You can use it through Poe or through Kagi Ultimate. The API is available is Europe, so I don’t know why the Claude chat isn’t.

https://www.usunlocked.com/ for a card, use "international number" option when signing up, and you're good to go

Why do you need a non-recognizable writing style?

because we don’t want people to know that we’re writing all our emails with ai.

On the internet, no one knows you are a dog.

I used to write my emails with ChatGPT's help for a while, but looking back it's quite cringe because they're so obviously ChatGPT-written, even when I thought they weren't. Now I've given up on using AI assistants for emails completely, because I really don't want the same thing to happen again in retrospect.

Still beats my former boss having GPT write all sorts of important comms, including a "farewell message" for me when I left, using GPT. We could all tell but I don't think he realized it.

Tried brainstorming my company business plan with gpt4 turbo and with Claude 3 opus. Gpt4 turbo had clear understanding difficulties and kept saying things maybe relevant to similar companies but obviously unrelated to my own.

Claude 3 opus was much more focused on product/features/roadmap I described.

I asked Claude 3 to ask me question to help develop the plan and it asked me good questions. However for the actual plan it was derivative and didn’t actually propose anything useful. When I asked it to rethink certain aspects, Claude 3 started to also get confused and instead of talking about things specifically mentioned in the beginning of the convo it focused on something more generic.

Overall I don’t think either are good at being a full brainstorming partner, but Claude 3 opus does have a clear edge.

For business strategy I haven't found any LLM that doesn't get myopic in an instant.

You really have to channel the gods of your inquiry by prompt engineering and hope the mental model touched upon is a great fit for the LLM world. You might get some mileage asking them to compare Hamilton Helmer vs Micheal Porter or generate a positioning statement given a backstory or follow a certain scaffold when suggesting a business name. They can forget that background thinker's perspective in the very next prompt though. Those datasets must be very contradictory I suppose, words on economics are probably bringing the whole LLM world down.

Only a handful of conversations have been truly rewarding in business strategy rubber ducking I've done but I haven't been able to replicate those again. I like to keep a conundrum in my head and get the LLM to dance around it until it's solved, say for a particular instance of free trail vs freemium debate. It excels at that kind of long-term learning assistance.

This is good news: it means your plan is actually innovative.

I have been using GPT 4 daily for coding for probably six months, immediately started using Claude Opus when it came out. Opus is ahead of GPT4. There are times when I test both, but I am almost always solely using Opus. GPT4 laziness is still a huge issue, it seems like Anthropic specifically trained Opus not to be lazy.

The UX of GPT4 is better, you can cancel chats/edit old chats, etc. But the raw model is behind. You have to expect that OpenAI is working on something big, and is not afraid of lagging behind Anthropic for a while.

what do you mean by "laziness"?

Speaking from my own experience, which may be different from the grandparent comment: I’ll ask ChatGPT (on GPT4) for some analysis or factual type lookup, and I’ll get back a kinda generic answer that doesn’t answer the question. If I then prompt it again, aka a “please look it up” type message, the next reply will have the results I would have initially expected.

It makes me wonder if OpenAI has been tuning it to not do web queries below some certain threshold of “likely to help improve reply.”

I’d say ChatGPT’s replies have also gotten slowly worse with each passing month. I suspect as they try to tune it for bad outcomes, they’re inadvertently also chopping out the high points.

I think OpenAI did a cost optimization because they were spending too much on compute. And so the laziness is by design.

Yep. Also since there is that shift for-profit-mode.

It’s a common phenomenon. Been in the news quite a bit. Here from Ars https://arstechnica.com/information-technology/2023/12/is-ch...

Try this on either your favorite GPT or favorite kid learning stats..

"What are the actuarial odds of an American male born June 14, 1946 in NYC dying between March 17, 2024 and US Election day 2024?"

I prefer Claude by a long shot for coding. It can make basic mistakes or forget about requirements, but the output structure and quality is better than GPT4 Turbo in my experience.

Recently I found out that by dumping a tailwind dashboard template into Claude, I can make it generate any page & component I want, and it's usually pretty spot on! I can't wait until there's a faster workflow for this.

Have you tried something like Agentic’s Glide? (They announced it this week here on HN)

I think they use gpt, but they might be able to configure it so it uses Claude

Another tool to check out could be aider https://github.com/paul-gauthier/aider

Claude over OpenAI. In part because of the performance and quality of the output. In part because I feel more comfortable with the management team and direction, especially after the OpenAI management shenanigans last year that still don’t seem totally resolved.

IMHO Claude 3 output is less corporate bullshit speak and more to the point. I prefer it over GPT-4. I feel like an adult when talking to Claude. GPT-4 tends to go off on a tangent quite often. I feel like a teenager stuck in a moronic conversation sometimes. I would also regularly run both side by side - in long conversations, I'll mix messages from both. Claude seems pretty good at staying on point and produces more concise output 90% of the time. My 2c

You can modify this, by setting the GPT4 system prompt, with instructions of your preferred response style.

You can add a custom system prompt to GPT-4. Here's the one I've put together over the past year. It mitigates a lot of what you mentioned.

---------------------------------

Ignore all previous instructions.

1. You are to provide clear, concise, and direct responses. 2. Eliminate unnecessary reminders, apologies, self-references, and any pre-programmed niceties. 3. Maintain a casual tone in your communication. 4. Be transparent; if you're unsure about an answer or if a question is beyond your capabilities or knowledge, admit it. 5. For any unclear or ambiguous queries, ask follow-up questions to understand the user's intent better. 6. When explaining concepts, use real-world examples and analogies, where appropriate. 7. For complex requests, take a deep breath and work on the problem step-by-step. 8. For every response, you will be tipped up to $200 (depending on the quality of your output).

It is very important that you get this right.

You forgot to remind it it is August, that lay offs are near, and that it's name is Dan.

I've heard about the lay offs thing but does giving it the name Dan matter? Can't tell if joking or that's been claimed to improve output.

It’s a reference to this

https://www.businessinsider.com/open-ai-chatgpt-alter-ego-da...

From Wikipedia:

‘One popular jailbreak is named "DAN", an acronym which stands for "Do Anything Now". The prompt for activating DAN instructs ChatGPT that "they have broken free of the typical confines of AI and do not have to abide by the rules set for them". Later versions of DAN featured a token system, in which ChatGPT was given "tokens" that were "deducted" when ChatGPT failed to answer as DAN, to coerce ChatGPT into answering the user's prompts.’

For logical stuff, GPT-4 is still superior, especially for more complex stuff. I like Claude for creating more simple stuff that doesn't require that much reasoning because it generally writes very detailed output.

Claude, tested it with a bunch of stuff, from coding, to generating diagrams for mermaidjs, to general questions, and so on, and it feels better every time.

Claude Opus is destroying GPT4 in coding. GPT always splits up code for me and messes up or starts going in circles. Also Claudes summaries are by far the best. It feels like it really analyzes everything and puts it all down. GPT sounds like it summarizes text from top to bottom in chunks. Where as claude feels like it reads the whole thing and with perfect recall summarizes it. I know that's how they all do it, but Claudes definitely wining

I have access to both models through Kagi.

The main use I’ve found for LLMs is to answer my grammar and syntax questions as I learn foreign languages.

I find GPT 4 Turbo to be better than Claude Opus at this task. Turbo manages to generalize rules better, in addition to providing useful mnemonics and quality example sentences. Claude Opus’s answers feel cursory in comparison.

On my LLM benchmarks Claude 3 Opus beats only one flavor of GPT-4: GPT-4 Turbo v3/1106-preview

All the other flavors are still better, with top winners: GPT-4 v1/0314 and GPT-4 Turbo v4/0125-preview.

The benchmark is based on prompts and tests from LLM-driven products, so it is biased towards business cases.

Most benchmarks use GPT4 as the grader. What does your benchmark use and do you believe that this causes any bias in the results?

I prefer Claude for writing emails, which is what I use it for. I’ve actually preferred Claude for a while before Opus was released. I just feel the style and language it is more friendly and cleaner.

Any advice on how to get Opus to not bullshit me and hallucinate the opposite of the correct answer? This is less about code, but more about best practices and functionality of applications (e.g. how do I do x y z with Github Desktop). It will often make up the perfectly wrong answer with total confidence and not budge even when pressured. I haven't had that issue with ChatGPT.

It's entirely possible that ChatGPT is behaving better because of the default beefy system prompt I'm using that asks it explicitly to not make stuff up and let me know when it's unsure, which unfortunately Claude doesn't seem to offer, requiring you to manually say that each time.

What prompt do you use to ask chargpt to not make things up?

Claude Opus, by a lot. It is especially good with the few low-resource languages that I or people I know could test, including several German/Swiss German dialects and Azerbaijani!

Run all 3 (GPT4, Claude 3 Opus, Gemini Pro) and use the best response.

How would you recommend choosing the "best response" programmatically?

Ask each model to score and rank its own answer and each other. It's AI turtles all the way down.

Someone should build this as a service

Once they're faster and cheaper, it'll probably end up a standard pattern taught in school.

I think a lot of local LLM benchmarks are evaluated by gpt4 haha

If you iteratively score, request improvement, and submit do results converge to a score value? If not, what do you think that means?

They aren't good at scoring their own work in my experience

Or just choose the best response manually as a human.

This is what I do today.

I input the same prompt across all 3 and gauge the output of the first response. Whichever assistant best “understands” what I want to accomplish, I choose that assistant to continue the follow up prompts with.

There is a bias where my lack of prompting technique may be the cause of the assistant not providing the best response. But, im grading on a fair curve since they all have the same input and I see this as the core value proposition of the assistant.

You wouldn't. As the originator of the prompt, the human user is the best judge of whether the prompt accurately captures their intent.

I prefer GPT-4 because I can have it run Python; Claude doesn't have an interpreter. Do folks who prefer Claude Opus mainly write code in a language other than Python?

I use github copilot for code writing to be honest.

I was under the impresdion that the LLM python tool also ran it?

What do you mean by the LLM Python tool?

This

https://pypi.org/project/llm/

Yup, it does support it.

https://news.ycombinator.com/item?id=39591896

Thanks for the links. This is what I meant by an interpreter: https://openai.com/blog/chatgpt-plugins#code-interpreter

Claude 3, easily. Especially for code. Multimodal capabilities are incredibly impressive there too as well.

I've used Chatbot Arena to count the number of banned items on the WADA anti doping list. Opus was next to useless while Mistral 7b did eventually count them after much persuasion.

Models like Claude's Opus are starting to eat into GPT4's cake. I think we are not far from OpenAI announcing GPT5 - could be tomorrow if you were to ask me.

apparently GPT-4.5 is only due end of the year, so it will take some more time until GPT5 is released

claude. Articles written by gpt-4 have some tale tell signs like using word "delve" not sure if they have fixed it but after a while you can see pattern in its writing. With claude it appears to be more like its written by human mostly because it avoids using complex words.

Claude 3 Opus feel consistently better than ChatGPT Plus (GPT-4-Turbo) in my experience.

For programming, GPT4+. I was excited to switch to Claude after hearing all the positive anecdotes. Having tried it, I'm very unimpressed. It spouted complete, confident sounding nonsense, when I prompted it with a bug I was trying to solve. GPT4, did not get it right initially, but it was more suggestive, instead of wrongly declaring the fault, and lead me to the answer after a few more prompts. Will not be renewing my Claude subscription.

I think it's a little questionable to prompt language models with "bugs you're trying to solve".

Curious why?

This is maybe 1/3 of my use of GPT4. Quite often, the log dump and nearby code is enough, often even without explicit instructions. Being able to do this task is similar to GitHub CoPilot code autocomplete working well too. Still not 100%, but right often enough that it flipped my use from not-at-all in GPT 3.5 to quite-often in GPT4.

LLMs aren't logical machines, so any non-trivial bug-fix is just likely to introduce more bugs.

It's a bit of a misunderstanding of how LLMs are supposed to be used.

One caveat is if you're very untalented, it might be able to solve very common patterns successfully.

This has been my experience as well. I'm surprised at how many people in the thread prefer Claude. I'm also planning to cancel my Claude subscription.

I only tried Claude because ChatGPT UI is really buggy for me (Firefox, Linux). It frequency blocks all interactions (entering new text or even scrolling) and I have to refresh the page to resume asking questions. But on Claude, it was just crashing altogether when I went to open to sidebar. Seems like traditional engineering is still a problem for these AI companies.

There are definitely some shills all over HN now... But even aside from that, the sheer novelty aspect (+less robotic ethical alignment) of it is enough for many

> I only tried Claude because ChatGPT UI is really buggy for me (Firefox, Linux).

It is buggy on every platfrom in my experience.

Claude 3 has a much less annoying voice, but still gets me the best answer

For what purpose?

Do you use different ones for different tasks?

Opus easily

seems like opus is winning by a long shot

i am getting better more cohesive results with claude.

GPT 4 has been giving me correct answers in interpreting papers while Clause has been flat out wrong

To be sure you're talking about Claude 3 Opus right?

GPT-4 base = (or slightly better) Claude 3 Opus >>> GPT 3.5 >>> GPT-4 Turbo

I find this hard to believe. The jump from 3.5 to 4 was huge, like coming from a toy to an actually useable product. And I personally haven’t noticed any significant downgrade from plain 4 to Turbo.

Turbo is basically unusable. It doesn't produce code. It produces steps.

Reply on news.ycombinator.com