A 14x cheaper AI model read a balance sheet on par with Fable
Reading figures out of two real annual reports, Claude Haiku returned the same balance-sheet numbers as the flagship Claude Fable 5, at fourteen cents a run against a dollar eighty-seven.
We maintain an open standard for recording every step an AI agent takes, in a format anyone can read, called the Agent Voyager Project (AVP). We used it to answer a question we kept hearing from people evaluating AI for financial work, which is whether a more expensive model reads a financial statement more accurately than a cheaper one. To test it we needed real filings with known-correct numbers, so we used a public benchmark of corporate annual reports paired with verified financial figures, called LEDGER, published by Artefactory under a Creative Commons CC-BY-4.0 license.
What we ran
We took two annual reports and one figure from each. From The Sherwin-Williams Company's 2017 report we asked for accounts receivable. From Aemetis, Inc.'s 2021 report we asked for accounts payable. Each report runs to roughly a hundred thousand tokens, so finding the right number means reading a long document, not answering from a snippet.
We gave the task to two Claude models. One was Claude Haiku, priced for low-cost work. The other was Claude Fable 5, the flagship. Both ran inside the same agent, the Claude Code agent, with the same prompt. The only thing that changed between the two runs was the model itself.
What we found
Both models returned the same numbers on both reports, and both read the figures correctly off the balance sheet. Claude Haiku answered for about fourteen cents per run. Claude Fable 5 answered for about a dollar eighty-seven, close to fourteen times more.
| Model | Price tier | Accounts receivable, Sherwin-Williams FY2017 | Accounts payable, Aemetis FY2021 | Cost per run |
|---|---|---|---|---|
| Claude Haiku | Low-cost | $2,104,555,000 | $16,415,000 | $0.14 |
| Claude Fable 5 | Flagship | $2,104,555,000 | $16,415,000 | $1.87 |
Both answers match the figures printed in the filings. The full run cost $4.02.
What the result suggests
For this task, the choice of model counted for far less than its price implied. That fits what we see in our own work, where the reliability of an agent depends mostly on how it is built rather than on which model runs inside it. In an earlier test we held the model fixed and changed the way the agent was constructed instead. Reading a page and rebuilding it as HTML, the weakest version scored 67 percent and the strongest 95 percent, on the same model throughout.
This is why we build pctx, the execution layer from Port of Context that sits between a model and the systems it reads from or acts on, so an agent can work on sensitive financial data in a controlled and auditable way. The model is one component inside that system, and rarely the one that decides whether the work can be trusted.
What comes next
We are running the same test across a larger set of filings and adding more models, and we will publish the fuller result on the Context Board.