Beane · built by Jatin Hans
I built Beane, a deep research harness. It reads around a hundred sources, checks every quote against the page it came from, and hands back a report with a numbered source behind each claim.
Official leaderboard, GPT-5.5 judge
232 research tasks across 22 fields, in English and Chinese, built from expert-written work. Judged by the benchmarks' maintainers and listed on 28 October 2026.
Also on these boards:
Official GPT-5.5 leaderboards. OpenAI Deep Research is listed as GPT-o3 Deep Research on DeepResearch Bench II. CellCog Max, first on DeepResearch Bench at 56.72, is not on DeepResearch Bench II.
Every entry on each official GPT-5.5 leaderboard, listed 28 October 2026, one dot each; DR is Deep Research. The bars come from the papers: the strongest systems tested in DeepResearch Bench II met under half its checks, and 50 is level with the reference report in DeepResearch Bench.
On both boards, against each company's best entry.
All nine entries.
| DeepResearch Bench II | Total | Facts recalled | Analysis | Presentation |
|---|---|---|---|---|
| Beane | 60.52 | 56.7 | 66.3 | 87.8 |
| SPIA DeepResearch | 58.97 | 54.7 | 66.0 | 88.0 |
| OpenAI o3 Deep Research | 43.12 | 38.2 | 46.3 | 83.6 |
| Gemini 3 Pro Deep Research | 41.52 | 36.7 | 44.0 | 84.9 |
| Perplexity Research | 33.57 | 29.1 | 36.0 | 73.5 |
The top five, nine more systems between 51.12 and 55.29, then the entries from Google, OpenAI, Perplexity and xAI.
| DeepResearch Bench | Overall | Comprehensive | Insight | Instructions | Readability |
|---|---|---|---|---|---|
| CellCog Max | 56.72 | 57.06 | 57.67 | 56.58 | 53.81 |
| Beane | 56.14 | 56.72 | 56.93 | 56.17 | 52.81 |
| ATH-Voicepica | 55.99 | 56.37 | 56.77 | 55.92 | 53.27 |
| SPIA DeepResearch V2 | 55.85 | 56.25 | 56.74 | 55.70 | 52.96 |
| OpenAI Deep Research | 47.84 | 48.05 | 46.69 | 49.29 | 47.62 |
Each stage gets the smallest model that does its job. A model picks which page is worth opening; code checks that every quote is on its page.
Python finds each quote on its page first. No quote, no finding.
Claims trace to a passage it read, not a search snippet.
A number in brackets means that page says it. Anything derived is marked.
Runs side by side, and each can pause and pick up again.
Figures from Beane's first 100-task DeepResearch Bench run, October 2026.
AI is restructuring labor markets primarily by reallocating tasks, changing skill demand and hiring, and redesigning work, not through an established, uniform collapse in employment. Whether displacement becomes net job loss depends on complementary human tasks, new work, organizational choices, and demand growth; effects therefore differ across industries and workers.[5][10][1][2]
Most of the questions I have been paid to answer were research questions, and the answer always sat across forty browser tabs, three PDFs and a regulator's spreadsheet.
So I drew a researcher that does the reading properly and proves every claim in code, and built it with Claude as my pair programmer. The name comes from Billy Beane of Moneyball, who spent only where it won games. I do the same with models.