Balyasny Says Codex Cut Economic Analysis From Two Days to 30 Minutes
Charlie Flanagan says Balyasny Asset Management’s internal AI platform has moved from a coding tool into a firmwide workflow system, with 97% of employees using it daily across investment research, software development and operations. He argues that GPT-5.5 and the Codex harness are shifting AI from systems that search to systems that do work, citing economic analysis compressed from two days to 30 minutes and earnings-report analysis moving closer to real time.

The operational measure is speed, not novelty
Charlie Flanagan describes Balyasny Asset Management’s internal AI platform as embedded across the firm, with practical value measured in elapsed time and breadth of use. His clearest quantitative statement is that “economic analysis that used to take two days is down to just 30 minutes.”
The deployment is similarly broad. Flanagan says 97% of the firm is now using the AI platform that Balyasny builds “on a daily basis.” He names three categories of work: investment research, coding, and automation of back-office operations and processes. The point is not that each employee is doing the same task with AI, but that the platform has moved beyond a specialist developer use case into multiple parts of the firm’s workflow.
That matters because the center of gravity is workflows. Flanagan’s example of two days compressed into 30 minutes is framed as economic analysis, not merely code generation. In the same minute, he places investment research, software development, finance operations, and back-office automation under the same platform umbrella. The benefit he describes is therefore not only faster programming; it is faster institutional work where analysis, code, operations, and presentation can sit closer together.
Codex is the mechanism for systems that do work
Flanagan puts the strongest label on Codex directly: “Codex has been a game changer.” The supporting statement is tied specifically to GPT-5.5. A text overlay states, “GPT-5.5 has unlocked a new level of intelligence.” Flanagan expands the point by saying the 5.5 model has unlocked “a new level of intelligence, but also a new level of work that the system is capable of doing.”
The 5.5 model has really unlocked a new level of intelligence, but also a new level of work that the system is capable of doing.
The important distinction is between intelligence and work. The platform is not described only as a better question-answering system. Flanagan says Balyasny first adopted it “from a coding perspective,” then quickly realized it was “capable of doing a lot more than coding.” Coding is presented as the entry point, not the limit of the system’s usefulness.
His broader formulation is that 2026 is “the year that we go from systems that can do search to systems that can do work, enabled by the Codex harness.” In that framing, the Codex harness is the layer Flanagan credits with moving the system from retrieving or summarizing information toward executing bounded tasks inside enterprise workflows.
The examples show work inside existing tools
The on-screen examples reinforce Flanagan’s point that Codex is being positioned for work beyond writing code. One displayed interface, attributed on screen as “Example content provided by OpenAI,” shows a data dashboard and slide-generation workflow for finance review material. The visible prompt asks a “Stratfin Slide Creator” to generate a new slide for “Paid Compute Margin” on the “Close Slides” tab of a dashboard. The screen references “Month-End Reviews,” margin analysis, and product-line views, including “68%, +3pts M/M and +5pts vs. Plan” and a date range from October 2025 to September 2026.
The most revealing text in that OpenAI-provided example describes the type of work being illustrated: “We’ve moved the manual finance assembly layer into Codex: It produces versioned dashboards, QAs against controls, drafts close slides, and turns trends into source-backed commentary.” The screenshot does not need to be read as documentary evidence of Balyasny’s own finance process. Its editorial function is narrower: it shows the kind of workflow OpenAI is using to illustrate Codex as a system that can assemble review artifacts, check against controls, draft slides, and produce commentary tied to sources.
A second interface shows a more conventional software-development use case. An AI coding assistant is shown working on a backlog item labeled APP-2437: adding accessible labels to previous and next question buttons in a Codex app navigation area, plus a focused regression test proving both controls have clear accessible names.
The assistant says it will trace the relevant panel and tests, patch the labels, add a narrow regression, and keep the work scoped to APP-2437. It reports searches and commands, identifies a relevant component file, notes that the icon-only navigation buttons need explicit names, and starts by writing a regression test expected to fail before wiring labels through react-intl. It also notes unrelated local changes already in the workspace and says it will leave those untouched.
This coding example shows the bounded form of work Flanagan is describing: an issue from a backlog, scoped changes, file references, test creation, and attention to unrelated changes. The system is not shown as a detached chatbot; it is shown operating inside a software task flow. The finance example and the engineering example differ in domain, but they point to the same operating pattern: Codex being used inside a process where the desired output is not just an answer, but a completed artifact or change.
For investment work, speed to insight is the use case
For investment work, Flanagan singles out earnings reports. “When an earnings report comes out, speed to insight is critical,” he says. The value proposition is temporal: the firm wants to shorten the interval between new information and usable understanding. Flanagan says the new Codex harness and GPT-5.5 allow that to happen “in almost real time.”
That use case follows the same premise as the back-office and software examples. Earnings reports create time-sensitive analysis demands; an internal AI platform becomes valuable if it can compress the work required to turn new disclosures into insight. Flanagan presents Codex and GPT-5.5 as enabling materially faster investment research workflows.
He also frames the work as unfinished. “We are at the frontier, and we’re all figuring out what these systems can enable next,” he says. That requires “not just the building of the tools, but the imagination of what we can build.” The emphasis is partly technical and partly organizational: the firm needs the platform, but it also needs people to identify where the platform can change workflows.
In that sense, the 97% adoption figure is not only a usage statistic. It supports the idea that the platform has enough reach across the firm for new applications to be found in different kinds of work. Flanagan names investment research, coding, and operations as areas where Balyasny is already applying it, and he frames 2026 as the shift from systems that search to systems that work.


