I want my AI inside DEVONthink 4 to help me with mainly three tasks:
Reformatting Markdown files (things like “fault-tolerant” conversion of lists to tables, headings with text to tables and vice versa)
Some general rewording of given paragraphs or lists in Markdown files plus extraction of some key terms out of PDFs and Markdown files.
Maybe writing some AppleScripts via the Script Assistant
As local model I use Gemma3 which is OK for some of these tasks. However external “big” chatbots like Claude or ChatGPT do often a better job when I try them for these tasks in the browser.
So for DEVONthink I am thinking of making the leap to buy credits for an API key of one of the big providers. (I have a Claude subscription, but I cannot make this work to get an API key.)
My question is: What external “big” provider (OpenAI, Anthropic or Google) has a good balance of cost and results with tasks like the ones listed above? What is the experience of other users with these external providers?
Why not just throw $5-10 at each and play with the results? After testing them, you’ll have a better idea of what suits your needs. For myself, I messed with various AIs and realized Claude was where I needed to go, so that’s where I went.
Or get an OpenRouter API key to easily test all of them first. Afterwards usage of the preferred provider is recommended as it’s a little bit cheaper, more reliable & faster than OpenRouter in my experience and as the native support is also more powerful.
From what I see here in the forum, time and money would be better spend on learning to write code.
AI needs publicly available code to to steal learn from, and the visibility of AS on the net is tiny compared with, eg, Python, C(++/#), JavaScript or even venerable Perl. Consequently, AI-generated code as seen here is at best terrible, at worst not working.
Agreed. That’s my experience as well: When I ask Claude Code to write something in Python it does a pretty good job (at least it is working most of the time), but when I ask it to write something in Scala it rarely works the first try! Same, or even worse, for Nu scripts.