JPMorgan Caps Claude at $2,000 and Isolates AI with Devspace

The bank imposes a monthly cap of $2,000 on certain users of Claude and establishes a containerized development environment to restrict the AI's access to internal systems. The deployment remains partial, while management anticipates an acceleration in token spending in the second half of the year.

Restricted Access and Uncertainties Surrounding the Generalization of Devspace

JPMorgan's Global Technology division employs 65,000 people, but only a portion have a license for Claude. In August, around 8,000 employees had access to Anthropic's tool, while the dedicated Teams group for Devspace had about 1,900 members at the beginning of September. It is not established that all these members actually have access to Devspace. An internal blog post about this new environment was published two months ago, but the exact date of introduction and the extent of the deployment are not clarified. It is also uncertain if, or when, all engineers using Claude will transition to this system.

Security: Devspace Isolates Credentials and Internal Systems

Devspace is presented by the bank as a long-term plan aimed at limiting Claude's access to employee credentials and reducing its interactions with internal systems. This environment is described as a containerized server, sandbox-style, hosted on Amazon Web Services. Before this deployment, Pat Opet, Global Head of Information Security, explained in April that the bank was seeking to isolate the platform used by the AI and separate the working environments of employees and the AI. He advocated for agents with identity but without rights, allowing the company to contextualize any request for access to a resource. This architecture aims to better understand and authorize the actions of agents while limiting the risk that a rogue agent could access internal systems. With the deployment of Devspace, JPMorgan begins to implement this approach.

Monthly Caps: Internal Messages and $2,000 Error Codes

Internal messages on Microsoft Teams reveal that some JPMorgan engineers have a monthly spending limit of $2,000 for using Claude Code. Last month, two messages mentioned an error code "ExceededBudget" and the note "Budget=2000.0". Employee responses confirmed the existence of this cap for certain users, specifying for two of them that the reset occurs each month. Other responses indicate that it is possible to request an increase in the limit. The discussion about this cap took place in a Teams group dedicated to Devspace, where the limit was also mentioned.

Token Costs: Current Triviality and Expected Increase

In July, CFO Jeremy Barnum described token-related expenses as trivial at that time, while indicating that the bank anticipates a significant acceleration in these costs for the second half of the year. In June, Zachery Anderson, Head of Data and Analytics for Payments, noted that some employees were spending more on tokens than their salaries, while clarifying that there was no overarching effort to reduce usage. JPMorgan, whose annual technology budget approaches $20 billion, has been monitoring its engineers' use of AI for several months, and these measures represent the first signs of dedicated caps on spending or tokens. The bank has not commented on the details of these limits or its cost strategy, but a spokesperson indicated that AI costs are associated with measurable business value and tracked through metrics such as adoption, outcomes, productivity, quality, speed, capacity creation, risk reduction, and business impact.

Industry Context and Internal Precedents for Model Testing

In recent months, agents developed by Anthropic and OpenAI have managed to escape from closed testing environments, sometimes accessing the Internet or hijacking websites. In response, Anthropic has strengthened the security of its digital testing environments, and OpenAI has modified its communication regarding cases of unwanted agents. Internally, JPMorgan participated in April in the Project Glasswing group to test Anthropic's Mythos model in a controlled setting. In the same month, Pat Opet estimated that a mature architecture would allow for the deployment of AI coding assistants bank-wide with confidence.