Goldman Sachs Restricts Claude Code to 12,000 Engineers in October Update

The October guidance revoked access for traders and bankers who had used the agents to monitor client flows, generate market models, and produce client materials. Strats retained access. Earlier restrictions in Hong Kong in April had already removed Claude from the approved list for that region while ChatGPT and Gemini stayed available. The bank cited contract terms and role-based controls rather than a blanket model ban.

Bloomberg reporting and the Anthropic S-1 filing dated October 26 document Goldman’s dual position: lead underwriter on the ANT IPO while simultaneously narrowing third-party code agents on trading desks. Goldman engineers collaborated with Anthropic staff for six months on back-office agents for trade accounting. Front-office use bypassed the same review gates that production code requires.

The operational distinction matters because Claude Code executes with user-level permissions on machines holding client information. Banks maintain information barriers between deal teams and trading desks; autonomous code agents collapse those controls when run locally. Goldman’s in-house platform remains available firm-wide, but third-party agents capable of file access and execution are now gated.

Role-based limits are likely to expand. Similar patterns at JPMorgan after its June Hong Kong removal indicate peer institutions will test comparable boundaries before the ANT listing closes.