For those interested in a sample of what Claude Sonnet 4.6 can do with regard to summarizing documents - this is an example fake medical record which Claude Sonnet 4.6 has summarized with links back to the “original” records. Impressive.
2 Likes
NOTE:
The upcoming DEVONthink 4.2.1 will support Claude Sonnet/Opus 4.6.
1 Like
The Claude Sonnet 4.6 report above cost about $15 to create
This report from the same simulated records was created by Qwen3 235B for under 15 cents:
2 Likes
How did you actually create the report and what kind of input did you provide?
Personally I’m a big fan of concise summaries 
but nonetheless I would prefer the Claude report in this case. The Qwen3 report feels somewhat sloppy, not only due to its broken layout.
A Chromebook can also do 95% of the things a Mac Studio can but is much more affordable. But nonetheless I’m eagerly awaiting the arrival of the first M5 Ultra powered Mac Studio Prices are secondary if you need the best tools for your work.
1 Like
I created it with a Typescript/React web app which uses Langchain’s Refine Chain. That lets me work with immensely long documents and document sets without having to split the PDFs. I used a 1-paragraph prompt explaining the various sections of the report that I want.
I have an alternate/earlier version of the app which summarizes each file individually and then does an overall summary of the summaries. But the refine chain handles that much simpler.
Personally I’m a big fan of concise summaries
I agree the Claude report is nicer. The cost difference is stunning though - that’s what makes me ponder which to use. Since I am using it basically as an overview index to find details in the original documents (never relying on the AI itself for content) what Qwen3 does may well be sufficient in many cases.
Not sure what I will decide for my own use - but in any event I thought it may be hepful for others to compare what they can do. It’s really quite a notable price difference which suggests to me more competition may be coming with AI.
I’m eagerly awaiting the arrival of the first M5 Ultra powered Mac Studio
Totally with you there - will be interesting in particular to see what local LLM models can run on such a computer.
So far no AI company is making money, therefore I doubt that these prices are sustainable.
Yes that is the big question
Or will we see less expensive data center models overseas just like iPhone manufacturing is less expensive overseas?