How Jump Trading is scaling quant research with ChatGPT

This is an OpenAI customer story, so it is a showcase, not an audit. Still, it is a notable claim from a serious trading firm. Lucas Baker, who leads LLM research and development at Jump Trading, says GPT-6 Astra has expanded what his team can hand to agents, from routine coding to quantitative studies that test new hypotheses. Where the work used to need frequent human steering, he says the team can now focus on defining a secure, monitored environment and clear goals and let the agents find their own way.

Baker describes tasks meant to run for days, pulling from many data sources and making judgment calls about what matters. The agents, he says, can stack small improvements on top of each other and redirect their own efforts against criteria set at the start. In a business where, as he puts it, predicting slightly better than a coin flip at scale is enough to win, that kind of throughput is valuable.

We take the capability claim seriously and the evidence lightly. No performance figures, error rates or trading results are disclosed here, which is normal for a vendor case study and also the reason not to treat it as proof. What we can evaluate is the design Baker describes, and on that, Jump sounds thoughtful.

The key idea is a wide sandbox with a hard gate. Agents can produce whatever they need inside a constrained, observable environment, and nothing moves forward without human review and acceptance. A trading signal an agent produces is treated like any other signal: usually informative, potentially wrong, and folded into a tightly controlled execution system. That is the right posture in a regulated industry where mistakes carry both financial and compliance consequences. Baker also notes that agents can be pointed at quality, security and monitoring, not just new features. We would like more companies to say that out loud.

The worry is where this heads. Baker's own forecast is a world of autoresearch, where loosely structured fleets of agents, coordinated by other agents, decide how to explore ideas, how to spend compute and which findings to keep, starting from little more than an open question. Humans would set the inputs, metrics and priorities. In a single firm with strong controls, that may be fine. Across an industry, it raises questions nobody in this story addresses. If many firms use similar models to search similar data, do their strategies start to converge? When an agent fleet finds an edge, can the humans signing off actually understand why it works, or are they approving outputs they cannot fully inspect?

That second question is the one I'm watching. Human review is only a safeguard if the reviewer has the time and understanding to say no. As agent runs stretch from hours to days and the outputs grow more complex, the review step can quietly turn into a rubber stamp. Our read: Jump's setup is a reasonable model for using powerful agents in high-stakes work, and the hard part ahead is keeping the human gate meaningful as the work behind it gets harder to see.

Jump Trading is a quantitative trading firm that builds predictive models from market data, news, events and alternative data across asset classes and time horizons. Baker frames recent progress as a series of step changes: early agents in 2024 struggling to write a single file cleanly, entire codebases in 2025, and in 2026 many agents collaborating on open research questions. Those characterizations are his, published by OpenAI.