Hi Everyone,
Adding LLMs to a Dash app is becoming easier and easier, but with so many AI services, it’s sometimes hard to choose one.
As we develop more community content and initiatives, we’re looking for community feedback on the topic to guide us.
Hi @adamschroeder Thanks for the poll!
We use Claude 3.7 Sonnet via LiteLLM, which exposes it as an OpenAI-compatible API. Should I select OpenAI or Claude? Thanks!
Good question, @hongbo-miao . I would say select Claude.
General experience from another forum:
The winner is here: ChatGPT (shortest code, least changes were necessary)
=> Also good results using: DeepSeek, Perplexity, Gemini, Mistral (Mistral is really giving source code output very fast)
=> Much less results: Grok3, Claude, Qwen, Llama
Each AI program was given this same prompt: ‘Create a xxxx program that calculates the first 1000 primes’
Note: No xxxx rules were supplied to begin with, only the prompt thus.
Only later some individual xxxx syntax rules were supplied where necessary.
ChatGPT 4o (Plus): very good: correct program (with minimal (=2) changes by user)
It’s not that deep 
I’ve been using OpenAI extensively, but lately, Grok3 has outperformed it with fewer errors and better handling of complex integrations. If Grok3 falls short, I’ll switch to OpenAI’s o3 Mini-high to see if it can debug where Grok3 failed, and vice versa.
I have had coworkers rave about ChatGPT, but it always leads them down a terrible path as the project develops out of the “toy” stage. I have tried to use both ChatGPT and Claude to get their projects back on track, but they both typically diverge from a useful solution.
Starting with a fresh prompt in Claude has typically come very close to a reasonable solution. Pasting the same prompt into GPT has never achieved results as good, and usually the code crashes or just does not do what you ask. Again, I am talking about real world problems, not benchmark coding problems or toy/tutorial projects.
For data security considerations, we utilize a privately deployed deepseek-r1 model along with Dash to develop various intelligent agent applications.
Happily used OpenAI’s ChatGPT, which tends to give well structured code that often works.
For a “learning experience”, I’ve also used Gemini, which tends to give buggy code. More often than not I would try and identify what is wrong with the suggested code. That in itself was useful, but I would hardly recommend it!
Have a setup using OpenWebUI deployment which has a similar interface to ChatGPT. You can add many different APIs to it including OpenAI, Anthropic or Gemini as well as some RAG integration. It’s great to be able to mix and match models as well as try out new ones from different companies as they are released.
For day to day tasks I often use chatgpt-4o-latest but recently I have been quite impressed with Claude 3.7 sonnet thinking LLM. Anecdotally at least it feels another level of intelligence higher than other models I’ve used for coding tasks.
Kinda biases the result when there’s no option for ‘None’. Why would I need a theft machine to write crappy Dash for me, I’m more than capable of doing that myself!
Github Copilot (GPT-4o)
OpenAI, typically 4-o. If the problem persists, o1 which will simulate the problem and provide additional debugging solutions within a response. This is for plotly.js. I don’t use Dash. GPT 4.5 is in research preview, and seems to fit between 4-o and o1. But the code doesn’t seem to be as good as either, if that makes sense.
I really like the structure and summary of ChatGPT, if I reach the free limit, Claude.ai is also good, I like the dark-themed code description on the right side.
I personally use OpenAI out of habit and because it was the first AI I used. But I find that the last version of Claude is better for programming questions on Plotly and Dash.
Interesting, I had no idea. Thank you for sharing.