September 2026 Benchmark

With Sleuth, AI tools find more

We asked Claude, ChatGPT, and Perplexity five business development and strategy questions, once on their own and once with Sleuth connected, then counted the programs, deals, or assets in each answer. Every answer with Sleuth listed more.

Hamed Ahmadi, Chief Technology Officer. September 22, 2026.

63%

more of the landscape on average, across 15 tests

40%

more at the median

15 of 15

tests found more with Sleuth connected

All 15 tests

Each cell is one AI tool answering one question, on its own and then with Sleuth connected. The number is how many more entities the answer listed.

All 15 tests: the lift in entities listed with Sleuth connected, for each AI tool and question
QuestionClaudeChatGPTPerplexityAll three
Scouting assetsChina-origin CLDN18.2 assets+57%60 to 94 assets+31%98 to 128 assets+111%35 to 74 assets+66%+57%averagemedian
Listing dealsI&I deals, $50M+ upfront+52%77 to 117 deals+32%117 to 155 deals+150%18 to 45 deals+78%+52%averagemedian
Market analysisBTK inhibitor share in CLL+33%21 to 28 entities+33%15 to 20 entities+13%8 to 9 entities+26%+33%averagemedian
WhitespaceOncology modalities, EGFR focus+40%10 to 14 entities+33%15 to 20 entities+70%10 to 17 entities+48%+40%averagemedian
LandscapeIL-23 landscape+93%42 to 81 programs+6%65 to 69 programs+195%19 to 56 programs+98%+93%averagemedian
All five questions+55%+52%averagemedian+27%+32%averagemedian+108%+111%averagemedian+63%+40%average of all 15median of all 15

What changed in the answers

Side by side, the answers with Sleuth also had the hard columns filled in, whole classes of drugs the AI tools missed on their own, and one row for each drug.

The hard columns, filled

Asked for I&I deals with $50M or more upfront, Claude on its own marked royalty terms “not disclosed” on 52% of its rows. With Sleuth connected, 89% of rows had real terms.

52%

of royalty-terms rows read “not disclosed” from Claude on its own

89%

of rows had real terms with Sleuth connected

A whole class it missed

Claude’s IL-23 landscape left out the TL1A × IL-23p19 bispecifics and the oral IL-23R drugs. With Sleuth connected, both were in.

One drug, one row

ChatGPT listed CPO102 and SYSA1801 as two drugs. They’re the same CSPC ADC, and with Sleuth they became one row.

Past the paywall

Perplexity said it couldn’t build a full deal list from public web search, and stopped at 18 deals. With Sleuth, it returned 45.

How we ran it

We asked the same five questions in each app, the way anyone would use it, once on its own and once with Sleuth connected.

Questions

Five, covering the work Sleuth users bring: scouting assets, listing deals, market analysis, whitespace, and a full landscape. Some listed every column to return; others were one line.

AI tools

Claude (Opus 5, high effort), ChatGPT (GPT-5.6 Sol, high effort), and Perplexity (Best), each in its own app with web search, not through an API.

Two settings

Each question once on its own, and once with Sleuth connected through its MCP connector. Five questions and three AI tools make 15 tests.

Sleuth’s part

The AI tools weren’t told whether or how to use Sleuth. Each wrote its own requests, called Sleuth one to six times, and chose what to keep.

What we counted

The unique entities in each answer: the main thing it lists, such as programs for IL-23 or deals for I&I. Answers were anonymized before counting.

The math

Each test’s lift is the change from alone to with Sleuth. 63% is the average of the 15 lifts and 40% is the median. Added up, the answers went from 610 entities to 927.

When

September 2026, on the versions available then.

What the test doesn’t show

We counted what each answer listed, not whether every entry was right.

We ran each test once per setting. AI tools answer a little differently every time, so any single result could change.

Two questions, market share and modalities, have short lists (8 to 28 entities), so a few entries make a big difference to the percentage.

We compared AI tools with and without Sleuth, not Sleuth with a database.

We ran it and counted the results ourselves, and it isn’t peer reviewed.

Why the answers changed

The challenge AI tools face in answering a question like the global IL-23 landscape is that the information is scattered across hundreds of data sources of uneven quality, behind different interfaces, and in several languages. Some assets barely appear in public (a preclinical candidate may be mentioned in a single publication or patent), while others appear many times under scientific names, brand names, trade names, and aliases.

Finding the data is only the first step. The insight comes from connecting the pieces: target to asset, asset to trial, trial to owner, owner to deal, deal to region. Resolving conflicts and duplicates takes fine-tuned algorithms, domain-specific classifiers, a lot of compute, and a rigorous quality control process. An AI tool has seconds to minutes, and a limited budget, to do that while it answers. So its answer depends on what it happened to retrieve, the model underneath, and how the question was worded.

A curated knowledge base like Sleuth’s removes that dependency. The entities are already resolved, the relationships already extracted, and every claim carries its source, date, and confidence score. Given structured, cited evidence, today’s models reason over it very well, and they do it consistently, whether you use Claude, ChatGPT, or Perplexity. The insight was already in the data Sleuth handed the AI tool.

Ask the same questions yourself

Copy a question into Claude, ChatGPT, or Copilot and see what comes back. Then connect Sleuth and ask again.

Scouting Assets

Identify all China-origin assets whose primary molecular target is CLDN18.2, across all modalities and stages.

Listing Deals

Find all business development deals announced from January 1, 2020 through today where the primary therapeutic area is immunology and inflammation and the disclosed upfront payment exceeds $50M.

Market Analysis

Model the market share and erosion path of BTK inhibitors in CLL over the next three years.

Whitespace

Which oncology modalities are primed to take over in the next five years, given market access dynamics, with a focus on EGFR targets?

Landscape

What is the current competitive landscape (including every program, all phases of development, actively developed assets only) for drugs targeting IL-23 or synonym targets and subtypes?

Set Up

Connect Sleuth to Claude, ChatGPT, or Copilot in three steps, then ask the same question again.

Ask your next question with Sleuth connected

Set up in Claude, ChatGPT, or Copilot