Goldman Sachs Locked Traders Out of Claude Code: Access Remains with 12,000 Engineers

⚡ The 5-Second Summary

• Goldman Sachs updated its internal rules in October: Claude Code from Anthropic and GitHub from Microsoft are now available to engineers only.
• Bankers and traders lost access, but strats — roles at the intersection of trading, quant research, and development — kept it.
• The bank's internal AI assistant, built on models from multiple vendors, remains open to all employees.

🔍 What Happened

In October, Goldman Sachs rewrote its internal policy: third-party agents that write and execute code on their own are now permitted only for the engineering division — roughly 12,000 people out of the bank's 47,000 employees. Claude Code and GitHub from Microsoft fell under the restriction. Traders and bankers used to rely on them: monitoring client flows, building models of market and position movements, and putting together text and graphic materials. The State of AI reports this, citing Bloomberg.

The mechanics are visible in the timeline. In July, the bank's CIO Marco Argenti told Bloomberg about "vibe coding" by non-engineering staff: a trader would first assemble a working prototype, then hand it off to engineers for refinement. The development queue shrank, but thousands of people without engineering training were running code on the same machines where client data resides. Then came the April story: Goldman removed Claude from its internal AI platform for bankers in Hong Kong, and FT wrote that the contract with Anthropic had been read literally there. ChatGPT and Gemini remained available in Hong Kong, and in June JPMorgan also removed Claude from its list.

🧠 Our Take

The reason this happened right now is stated dryly in the source: "a productivity and risk review." Productivity is about what was promised in July; risk is about the fact that no one was measuring anything. Our assessment: the bank didn't hit a wall over model quality — it hit the boundary of accountability. When a trader writes an agent themselves that has access to client data, audit and compliance can no longer tell who the author of a decision is — the human, the model, or a chain of prompts.

What this changes for the market: "vibe coding" has received its first clear corporate pushback from an organization that isn't afraid of experiments. Not a ban on AI — a ban on the opaque layer between employee and code. Based on experience, such a decision is usually followed by a separate class of internal tools: a sandbox, prompt logging, review of generated code. Someone will make money on that, and it won't be Anthropic.

Who wins and who loses. Goldman's engineers win: the monopoly on code has returned to them, along with a queue of internal customers. Traders lose on speed, and the source directly says some of them are annoyed. Anthropic formally loses nothing: the bank remains a partner, its engineers spent half a year building Claude-based agents for trade accounting and client onboarding, and Goldman, together with Morgan Stanley, is a lead underwriter of Anthropic's upcoming IPO. The question is how long the bank can keep selling investors the AI story while restricting that AI to twelve thousand of its own developers.

⚠️ What We Don't Know

The source retells Bloomberg, and Bloomberg itself relies on "people familiar with the matter" — there are no figures on how many people lost access or on specific incidents. It's unknown whether there was any data leak or agent error at all, or whether this is a preemptive decision. Separately: the details about Anthropic's IPO (listing timeline, valuation) in this piece look weakly substantiated — verify them against primary documents, not secondhand accounts.

Forecast: within a quarter, at least one more major investment bank from the top 20 will announce a similar restriction — access to third-party coding agents by role, not by division.

Disclaimer: not officially confirmed in detail; a single source — Bloomberg, as retold.

Source: The State of AI