17%→54%
The best model with web vs. with Aiera (3.2×)
The same frontier model, asked the same research question, once with web search, once connected to Aiera's MCP server. On the questions the open web can't answer, the gap is the whole story.
Everything is held constant — same model, same question, same scoring — except whether the model can reach Aiera. The delta is Aiera's contribution.
What a sophisticated user does today: a model with its own live web search. Clears the bar on widely-covered companies; falls short where the answer isn't on the open web.
The same model, connected to Aiera's MCP server (transcripts, filings, expert insights, and proprietary research). It retrieves the actual source and answers from it.
Score = share of an answer's required facts that are correct, graded by a held-out judge. Split by whether the open web could plausibly answer the question.
Two real questions from the benchmark, the same model, both conditions.
Built to be defensible, not flattering.
Every question's answer is grounded in a real Aiera source (transcript, filing, or research note); each required fact traces to a verbatim excerpt.
Questions are phrased the way an analyst asks them (but never name the source) so both conditions get a fair attempt.
A neutral frontier model, held out of the results, grades each answer fact-by-fact, plus a holistic cross-check.
The dataset stays private, only scores are published. Every model reaches the same MCP server.