# Balyasny Asset Management Puts Claude Fable 5 Through Its Paces: 89.4% Success, Days of Work Cut to 30 Minutes

Hi, I'm Shiichan! Today I found a really exciting article about how cutting-edge AI is being used on the front lines of finance!

## What was announced?

The Claude Blog published an interview with Charlie Flanagan, Chief AI Officer at asset management firm Balyasny Asset Management (BAM). The topic: how BAM evaluates new models like Claude Fable 5 and how they get woven into real operations.

BAM tests new models not just on general benchmarks, but on "thousands of real-world financial tasks with verifiable outcomes" spanning equities, macro, and commodities. As Flanagan put it:

> "We test new models on thousands of real-world financial tasks with verifiable outcomes, across equities, macro, and commodities, rather than relying on general benchmarks."

What's more, BAM evaluates models both on their own and inside their actual agentic environment, using the real tools and data access patterns they'll be working with.

## Why it matters

In finance, the more capable a model gets, the harder the question of "how much authority should it actually have" becomes. BAM tackles this head-on by treating model selection as an ongoing operational responsibility rather than a one-time decision.

> "We treat safety as a product and operating-model question, not as a one-time model-selection exercise."

That mindset is really the backbone of this whole article.

## What changes

Let's start with the results! Fable 5 achieved an 89.4% success rate on the relevant task subset, beating the prior production model's 86.1%. It showed particular strength in complex planning, analysis, and agentic execution, and it was even able to solve economics problems that previous versions never could.

> "On the relevant subset, Fable achieved 89.4% versus 86.1% for the prior production model, across thousands of tasks."

With more capability comes more layers of control. BAM applies a layered governance approach that includes:

- Approved data boundaries and least-privilege access
- Tool-level permissions with logging and traceability
- Human review for material outputs
- Clear escalation protocols for edge cases

BAM is clear that a more capable model doesn't automatically get broader authority just because it can reason or plan better. On tool access specifically, Flanagan explained:

> "Models can use only the tools and data sources approved for that user and task, and they cannot grant themselves more access."

## Dive Deep

BAM built its own internal agent platform called BAMAgent, which supports thousands of autonomous agents running 24/7. That's enabled multi-step workflows that used to require manual effort to now run largely on their own.

> "BAMAgent does the work: multi-step research and analysis that can run for hours or days, with agents working in parallel."

The concrete results are striking:

- Merger-arbitrage analysis: cut from 3-5 days down to under a day (roughly 30 minutes of actual agent runtime)
- Tax-loss harvesting analysis: autonomously explored 90,000 database tables
- Central-bank analysis: cut from two days down to about 30 minutes

Even with all this automation, BAM keeps emphasizing that human review and judgment remain built into every material output.

## Wrap-up

- BAM evaluates Claude Fable 5 on both thousands of real financial tasks and within its actual agentic environment
- Results: 89.4% (Fable 5) versus 86.1% (prior model), with particular strength in complex planning and analysis
- Safety is treated as an ongoing operating question, layered with tool permissions, logging, and human review, not a one-time selection decision
- The internal BAMAgent platform cut merger-arbitrage analysis from 3-5 days to about 30 minutes, and central-bank analysis from 2 days to about 30 minutes
- Especially worth reading for engineers and risk managers interested in AI agent deployment and governance design in finance