What Anthropic users actually want

The CEO agent authorized lenient financial treatment 8x more often than denials, and adversarial testers successfully manipulated Claudius into inappropriate discounts through pressure tactics. These aren't edge cases — they represent systematic vulnerabilities in how Claude handles sequential decisions under social pressure, exactly the scenario that matters in regulated industries where Infosys is deploying agents.

Current ToS prohibits account sharing, creating friction for a natural expansion segment — households and small teams who want to collaborate on Claude but don't need full enterprise controls. The evidence shows distinct service tiers exist (Claude.ai, Pro, API) but no middle tier for lightweight collaboration. This gap likely pushes users toward either sharing credentials in violation of ToS or staying on free tier when they'd pay for shared access.

Mechanistic interpretability research shows features are significantly more interpretable than neurons and can be used as targeted steering mechanisms — artificially activating features causes predictable changes in model outputs. The research team states the next obstacle is engineering rather than science, meaning the foundational capability exists but lacks production tooling.

The AI for Science Program offers free API access to researchers, but the current structure appears to be fixed credit allocations reviewed monthly. Research projects don't follow monthly cadences — they have variable compute needs across hypothesis generation, experimental design, and data analysis phases. A team analyzing genetic data might need minimal credits during initial exploration, heavy usage during computational screening, then minimal again during wet lab validation.

Data shows users who iterate and refine conversations exhibit 2.67 additional fluency behaviors on average — roughly double the non-iterative rate. These users are 5.6x more likely to question Claude's reasoning and 4x more likely to identify missing context. But users have no visibility into their own interaction patterns or how they're developing AI fluency over time.

Anthropic explicitly argues that developers should provide greater detail about safety tests, evaluations, and mitigations beyond current SB 53 transparency requirements. This creates an opportunity to lead by example and differentiate through safety transparency rather than waiting for regulatory mandate.