Jane Street details near-firmwide Claude use and research gains in Anthropic case study

Researchers at Jane Street now supervise teams of Claude agents that analyze code, monitor training runs and process results, according to an Anthropic case study. The trading firm describes deployment across nearly the entire organization, with longer-running agents helping researchers complete more studies rather than requiring constant attention at every step.

From interrupted tasks to weekend experiments

Jane Street builds models, strategies and systems for pricing and trading financial instruments. Its research depends on working through large datasets, but its early attempts to build agents that could complete research tasks on their own were inconsistent. Developers struggled to make those systems carry work through to completion, the case study says.

The firm identifies Claude 3.7 Sonnet as a turning point. After adopting that model, it saw agents handle long sequences of work without stopping midway or needing rework. Successive model generations then encouraged wider use across research, engineering and design. The account describes that progression as a change in reliability, not simply faster answers to individual questions.

The resulting workflow puts researchers in charge of several agents rather than every intermediate action. Agents examine code, watch training runs and process the results. Experiments that previously demanded continual supervision can now run over a weekend, Jane Street says. The claimed benefit is more room to pursue studies and test ideas while those tasks continue without constant intervention.

Two different measures of time saved

One researcher estimates completing about 50% more analysis than a year earlier. That is an individual estimate, not a firm-wide productivity measurement. The case study connects the improvement to getting more studies done and testing them faster. These are customer-reported results published by the firm’s AI supplier, not independently verified findings.

A separate example concerns a large Python application built by a researcher. Using Claude Fable, he first reorganized its sprawling code into a cleaner structure, then migrated it to another language in about a week. The case study says comparable work would normally take a team months. That comparison uses an estimated conventional workload, rather than a parallel team performing the same migration.

We’ve always had more ideas than we’ve had the bandwidth to execute on. Agents like Claude have dramatically improved our velocity, multiplying the impact each person can have.

Craig Falls, Head of Quantitative Research at Jane Street

A harder coding test, separate from productivity

Jane Street also evaluates agents against work its engineers have already completed, grading quality, accuracy and trustworthiness. Models that once passed just 1% of those tests eventually outgrew them, prompting the firm to build a harder agentic coding evaluation—a test of an AI system carrying out coding tasks. On that evaluation, Claude Fable 5.1 scored 32% higher than Fable 5.

Falls says Fable 5.1 solves more of the firm’s coding problems than either Fable 5 or Opus 5 in internal benchmarks. He also says its work remains readable over long, multi-step tasks, unlike earlier models that became harder to follow. Those observations concern success on Jane Street’s problems and the ability to follow an agent’s work, distinct from the researcher’s estimate of increased analysis.

The change also reaches beyond researchers and engineers. Designers now often write code for features that the case study describes as production-ready, then pass it directly to developers to merge. In that workflow, the handoff is no longer just a feature idea or design: it includes the implementation.