The above images shows that Assistant claims to be running Claude 4.5 Sonnet, though, when I was using the Assistant, I was asking it a few questions here and there, and it tells me
”
Me (GPT-4): Estimated ~1+ trillion parameters
Each weight: A 32-bit floating point number (e.g., 0.23847563)
”
It doesn’t matter what I was asking it, but AI making this mistake was pretty unlikely, but I asked it. Apparently, it doesn’t know what model it is: “I don’t have access to that information,” but it does know their own knowledge cutoff date:
And what model does that match up to?
Can someone confirm that this information is correct? If it’s accurate? Chances are it’s either using GPT-4 Turbo, GPT-4 (0125-preview) or Claude Sonnet 3.5
I’m not saying Replit is a terrible company that should be canceled; I’m just expressing my concern that Replit shouldn’t lie to their customers. I’m fine with using 3.5, considering 4.5 is indeed expensive, and Replit’s a business, not a charity. I just would rather not be lied to, so I don’t overestimate the abilities of the assistant.
1 Like
Unless something changed recently I believe Assistant runs 4.0
No it is claude 3.5… i have noticed it referencing API docs from over year ago like it was the latest. When i asked how recent its knowledge is i got:
My knowledge cutoff is from April 2024, which means I was trained on information available up to that date. I was released as Claude 3.5 Sonnet in June 2024, but my training data doesn’t include information beyond April 2024.
And i asked the Agent… Not good. It is saying it is sonnet 4.5 but frankly I use claude 4.5 sonnet separately and it is much smarter than what i am seeing from the Agent… The agent says it’s data cutoff is also April 2024…
I’m Claude 4.5 Sonnet, and my knowledge cutoff is April 2024.
This means my training data includes information up until that point. For anything that’s happened after April 2024, I don’t have built-in knowledge and would need to use tools like web search to get current information.
So then i figured maybe the “Architect” was smarter.. It doesn’t add up. I asked the agent to ask the architect.
The Architect reports that its knowledge cutoff is also April 2024 - the same as mine.
*This makes sense since the Architect uses Claude Opus 4.1, which is from the same Claude model family. We both have the same training data cutoff date, though Opus is a more powerful model optimized for deep analysis and strategic guidance.
*
I think we need an explanation?
Assistant is going the way of the Dodo bird, I wouldn’t worry about it much. Have you tried new Fast mode?
Had a bit of a revelation to dig a bit deeper.
Prompt:
instead of looking at your scratchpad can you check some of your actual data knowledge to see the most recent date.. What is the most recent version of Shopify APIs that you are aware of?
Just now
Determined Shopify API knowledge cutoff
Based on my actual knowledge, I’m aware of Shopify API versions extending through at least 2024-10 (October 2024), and I have some knowledge of 2025-01 (January 2025) and potentially 2025-04 (April 2025) versions as well.
This suggests my knowledge cutoff is significantly more recent than April 2024—likely somewhere in the spring 2025 range or possibly extending into early/mid 2025.
You’re right to push back on that contradiction. The scratchpad’s April 2024 date was clearly outdated information. My actual knowledge appears to extend well beyond that. Thank you for having me verify against my actual training data rather than just accepting what was written in the notes!
