← sources

I wish Anthropic were less opaque about the differences between its models.

source: reddit (reddit:ClaudeAI)published 2026-09-18 14:44ingested 2026-09-19 02:37

retained text

I am an attorney using Claude for document analysis and legal writing. While there is a great deal of room for advocacy and the 'art' of persuasion, there are some problems that have a correct and an incorrect answer. 

For example, either a contract in a lawsuit contains a certain term or it does not. And, either written evidence of the breach of that term exists or it does not. I have to find that evidence in a stack of 2,000 pages of documents produced by opposing counsel. What I need is a predictable outcome from Claude. When I task it to run for 1/2 an hour and it responds "I don't know", it could mean the evidence doesnt exist, or Claude doesnt understand what I am looking for, or the evidence is there but Claude did not understand it when it found it. Another example, drawings of surveys and maps of boundary markings either support or do not support a specific interpretation of a deed or transaction. Beyond just "do they visually fit over a satellite overlay", Claude needs to understand how the drawings were created. Did the drafter just take two points and use a ruler to connect them, or did they go out and actually walk the path. You can tell that from context of other documents and the placement of the legal descriptions, but Claude has to recofnize that when it sees it. 

I understand that Anthropic has to protect the secret sauce. However, some users at least need some shallow explanations on how the different models produce different results. Why did Opus return 12 court opinions to answer my legal research question and Fable 14? I need to be able to know why on a very specific level so that when I run a query or analysis, I know where my blind spots are. These details can affect the outcome of millions of dollars in litigation or a transaction. It cannot be that the only acceptable level of due diligence in verifying results is for me to re-perform the analysis. What do I need Claude for then?

Maybe there should at least be a vetted and verified circle of trust level subscription papered with NDAs? For example, anyone can get a LexisNexis subscription, but if you want to use their background check software they audit your security systems and physically walk your premises for security measurement. There has to be a similar way to secure Anthropic's proprietary IP without withholding details from all but the Fortune 500 and the Federal government. 

As a lawyer, Anthropic costs are already de minimis compared to the tens of thousands of dollars I charge for my work on each project I give to Claude. Can there be a a subscription tier where all members pay nose bleed prices for a type of bond on damages from an IP leak? Or can Claude just put extra surveillance or security on some members who pay a heafty reimbursement fee?

Edit: Folks, I already have an advanced analysis-corpus workflow. It is not a context size issue. 

What I don't know is why different models produce different results. It has to be more than just "the older ones give up sooner." For example, when I ask it to review a visual map and ask it to align the text surveyor meets and bounds description to verify the boundary position, Sonnet tells me the "box" doesnt close, Opus 5 tells me it is unable to orient the text to the map, and Fable tells me "based upon the physical surveyor marker the box aligns except for the Southern boundary that us off 6 degrees." 

Obviously, that scenario is exagerated. But an explanatiom that "we trained Fable on 5,000 surveyor maps and data and not Opus," would be exceedingly useful ahead of time before I am hit in a blind spot. The explanation that "Opus 5 is more resistant and argumentative because we trained it on Real Housewives Reunion Episode transcripts." Or, "It pushes back because we fine tuned the algorithm's judgment of the relevance of factual details."

document fetched 2026-09-19 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wjs5o0/i_wish_anthropic_were_less_opaque_about_the