# JPMorgan Sets $2,000 Claude Code Limit for Some Engineers

By John Kim

Canonical URL: https://www.tokenpost.com/news/business/37973
Published: 2026-10-22T14:44:07.000Z
Updated: 2026-10-22T14:44:07.000Z
Section: Business

> The cap applies to some users as CFO Jeremy Barnum calls an earlier $700 million annual AI expense estimate rough and stale.

JPMorgan Chase has set a $2,000 monthly Claude Code limit for some engineers as banks weigh AI productivity gains against costs they cannot yet measure consistently.

The limit resets monthly, and engineers can request a higher amount. It applies to some users, not the entire bank.

On JPMorgan’s Oct. 13 earnings call, Chief Financial Officer Jeremy Barnum described an earlier estimate of about $700 million in annual AI expense as rough and stale. He said the bank’s AI spending is spread across the institution and cannot be cleanly separated from other costs. Earlier estimates of AI benefits also included machine-learning tools such as consumer fraud prevention.

“The benefits aren't observable in our current quarterly earnings,” Barnum said of newer generative and agentic AI. Barnum also said token expenses could absorb part of the productivity gains.

Citigroup CEO Jane Fraser gave different measures of AI-related efficiency on the bank’s Oct. 13 earnings call. She said 14,000 engineers use Citi’s Devin coding tool, product launches were 30% to 40% faster, and technology vendor spending fell 30% year over year as more work moved in-house. Fraser said Citi was “launching new products at lower cost and faster speed.”

Those figures track launch speed and technology vendor spending, while JPMorgan’s estimate covers broader AI expenses. They do not provide a common basis for comparing the banks’ total AI costs or net returns.

Across 50 banks, 93 new AI use cases were announced in the second quarter of 2026, up 45% from the prior quarter. Twenty-seven percent of the new cases disclosed an outcome, compared with 41% in the first quarter.

JPMorgan’s AI strategy includes proprietary models, open-weight and open-source models, and some models running in its own data center. CEO Jamie Dimon described the goal of model routing as finding the “best, quickest, cheapest” option. The disclosed measures leave banks without a shared way to compare AI spending with the results it produces.