Claude Sonnet 5 Users Report Refusals, Pushback, and Memory Issues

Claude Sonnet 5 launched last week, but Anthropic’s newest model is already facing criticism from users who say it feels less natural, more argumentative, and harder to control.

According to complaints shared by users on Reddit, the model appears to struggle with some basic conversational behavior despite its expected benchmark improvements.

Users Report System Prompt and Memory Issues in Claude Sonnet 5

One of the biggest complaints involves Claude Sonnet 5 referencing its own system prompt or context memory during conversations.

Instead of silently following hidden instructions, users say the model sometimes openly explains them. This can make replies feel unnatural and distracting, especially when the user simply wants a direct answer or completed task.

Some users also claim the model exposes parts of its instruction-handling process, which breaks the flow of normal conversations.

Claude Sonnet 5 Accused of Refusing Basic Requests

Another major complaint centers on refusals and resistance.

Several Reddit users claim Sonnet 5 pushes back against simple commands, corrects users too aggressively, or creates disagreement where none exists.

Users describe the model as overly argumentative in some cases. Instead of completing a task, it may challenge information the user just provided or build a strawman argument around the request.

The issue appears linked to Anthropic’s attempt to make the model less overly agreeable. However, users argue that Sonnet 5 may now overcorrect in the opposite direction.

Some complaints go further. One Reddit user claimed Sonnet 5 refused to help with a basic task and instead accused them of fraud.

Excessive Correction Makes Tasks Harder

The main frustration is not just that Sonnet 5 disagrees. Users say it disagrees too often and at the wrong time.

Normal work requests can turn into long exchanges where the model focuses on caution, correction, or pushback instead of helping.

That makes the model feel less productive for users who want writing, coding, research, or routine assistance without unnecessary debate.

Users Suspect Sub-Agent Routing Problems

Some users also suspect Claude Sonnet 5 routes tasks through multiple sub-agents.

They believe these sub-agents may underperform or produce weaker answers, even while using more tokens. If that is true, users may end up paying more for responses that feel less useful.

Anthropic has not confirmed that this is the cause of the complaints.

Context Loss and Disclaimer-Heavy Answers Also Reported

Other users report that Sonnet 5 can lose context after only a few prompts.

Some also say the model gives too many disclaimers, hedges too often, or becomes overly cautious even when the request does not require that level of warning.

These issues add to the broader complaint that Claude Sonnet 5 may feel technically advanced but less smooth in real conversations.

Claude Sonnet 5 Faces Rough Early Feedback

Claude Sonnet 5 may still perform strongly on benchmarks, but early user feedback has been rough.

The complaints suggest that Anthropic may need to adjust how the model handles instruction-following, safety pushback, context memory, and disagreement.

This is not the only recent Claude model facing criticism. Users have also complained that Claude Fable 5 feels severely limited after being restored to the public.

Claude Fable 5 will remain limited behind Anthropic’s pay-as-you-go plan for now, but that could change once the company has enough capacity.

Via Neowin