Leaderboard

How many reported figures an AI-built model gets wrong: the same 50 companies (35 US, 15 European), one DCF and one comps table each, built by each model with web search alone, with deskstack's data, and with deskstack plus its reviewers. Every figure the scorer called wrong was then read by hand against the filing. Models so far: 1. More models soon: each run is a paid run, approved before it starts.

DCF, 50 companies

Model	Setup	Models finished	Wrong reported figures	Cost per model	Time per model	Source
Claude Sonnet 5.5	A: web search only	50 of 50	20 of 573 (3.5%)	$0.23	58 s	analysis
Claude Sonnet 5.5	B: deskstack data	49 of 50	0 of 593 (0.0%)	$0.07	18 s	analysis
Claude Sonnet 5.5	C: deskstack + associate + VP review + fix	50 of 50	0 of 606 (0.0%)	$0.10	32 s	analysis

Comps table, 50 companies

Model	Setup	Models finished	Wrong reported figures	Cost per model	Time per model	Source
Claude Sonnet 5.5	A: web search only	45 of 50	16 of 445 (3.6%)	$0.56	171 s	analysis
Claude Sonnet 5.5	B: deskstack data	50 of 50	0 of 1,021 (0.0%)	$0.17	36 s	analysis
Claude Sonnet 5.5	C: deskstack + associate + VP review + fix	50 of 50	0 of 1,010 (0.0%)	$0.22	51 s	analysis

How the figures are checked

Each submitted figure is scored against the filing it should come from, then every one the scorer calls wrong is read by hand and classified. The analysis of each run lists every case.
The setups with deskstack read the data they are scored against, so their 0 is partly by construction. deskstack's data is checked separately: a golden set of questions with known answers on every push, and a nightly re-read of each answer from the filing itself with independent parsers, edgartools for SEC filings and Arelle for European reports.
Generated from bench/results by server/bench/leaderboard.ts. The runner, the frozen company set and the method are in bench/.

Independent open-source project. Not affiliated with or endorsed by Anthropic. Not investment advice.