JPMorgan has introduced JPMorgan Claude Code spending limits of $2,000 per month for certain engineers, internal Microsoft Teams messages seen by Business Insider show, marking the first confirmed indication that the bank is capping AI token expenditure after months of carefully tracking usage across its technology workforce.

The cap sits alongside a parallel security initiative called Devspace, a containerised, sandbox-like coding environment hosted on Amazon Web Services. Together, the two measures signal that the bank is moving from passive monitoring to active governance of what has become a rapidly expanding cost line.

How the JPMorgan Claude Code spending limits work

The messages, exchanged in a Teams group for engineers with access to Anthropic‘s Claude, flagged a new error code: “ExceededBudget” and “Budget=2000.0.” Responses from employees confirmed a monthly ceiling of $2,000 for certain engineers using Claude Code, with costs resetting each month. Users can request an increase to their individual limit, according to replies in the same thread.

JPMorgan, which carries a nearly $20 billion annual technology budget, had tracked engineers’ AI usage for months but had not previously moved to set explicit spending or token caps. A token is the basic unit of text that a large language model processes; token costs accumulate quickly at scale, and the issue has become increasingly pointed across Wall Street.

In June, Zachery Anderson, chief data and analytics officer in payments at JPMorgan, said at a tech event that some employees are “spending more on tokens than their salary,” though he added there was no firmwide effort to scale back usage, according to Semafor. During the bank’s second-quarter earnings call in July, Chief Financial Officer Jeremy Barnum described token expenses as “trivial” at that point but said the bank is “forecasting some meaningful acceleration in that number for the second half of the year.”

The $2,000 monthly figure sits against a backdrop of rising published benchmarks from Anthropic itself. On 15 April, Anthropic more than doubled its public estimates for Claude Code token costs, according to Business Insider. Archived versions of Anthropic’s guidance showed that before 16 April, the previous figures were $6 per developer per active day, with costs remaining below $12 per active day for 90% of users. Business Insider reported that the revised estimates put the average cost at $13 per developer per active day, with costs remaining below $30 per active day for 90% of users. That doubling of Anthropic’s own benchmarks illustrates why a $2,000 monthly ceiling can be reached far faster than it might initially appear.

Devspace: isolating AI agents from the enterprise

The spending cap was mentioned in a Teams group centred on a new coding architecture called Devspace. Until recently, some developers ran Claude directly on their desktops, giving the model proximity to employees’ credentials and internal systems. Devspace changes that arrangement by restricting Claude’s access to those credentials and limiting its ability to interact with internal infrastructure, according to a JPMorgan spokesperson.

The architecture is described in an internal blog post, published roughly two months ago, as a containerised sandbox hosted in AWS. A sandbox, in software terms, is an isolated environment where code can execute without affecting the broader system, an arrangement particularly suited to running AI agents, which can make decisions and carry out tasks autonomously.

Pat Opet, JPMorgan’s global chief information security officer, outlined the rationale for this approach at a cybersecurity conference fireside chat in April, before Devspace was introduced. He said it is risky to place powerful AI tools on the same desktops that hold employees’ emails, sensitive files and personal information. His stated aim was to build an architecture that “isolates the platform that the AI uses” and creates distinct environments for employees and AI systems.

“Ideally, we would want these agents to run in an ecosystem where they have no entitlements. They have an identity, but no entitlements,” Opet said. “When these tools need access to a resource, we want to be in control of understanding the context around that.” Under that model, if an agent escaped its environment, it would lack the credentials needed to access internal systems or, as Opet put it, “abuse the enterprise.”

As of now, only a portion of the bank’s 65,000-person Global Technology division uses Claude. An internal dashboard viewed in August showed around 8,000 people held Claude licences, while the Devspace Teams group had approximately 1,900 members in early September. The spokesperson confirmed Devspace is relatively new, with new features continuing to roll out; it remains unclear how many engineers have migrated to the new environment or on what timeline the broader rollout will proceed. Opet said in April that successfully building this architecture would generate the “trust and confidence” needed to scale AI coding assistants across the firm.

Share.