Mento VC's SAB: AI Aces Startup Research, Fails Growth Prediction

Can AI be a VC analyst? Venture firm Mento VC's answer: great at the work, bad at fortune-telling.

Mento has published its Startup Agent Benchmark (SAB), which as of October compares 14 models from four vendors including OpenAI, Anthropic and Google. The research half has 960 tasks across six analyst workflows—product and customer, market and why-now, competition, traction, sourcing screens and more—split between synthetic companies and a frozen snapshot of Mento's real company data. Answers are graded on 4,560 criteria, and every field allows "unknown" when evidence is missing.

Research scores are strong. Claude Opus 5.5 leads with 97.2, though its margin over the runner-up isn't statistically significant. Sourcing and product tasks are near the ceiling for every model; market sizing and competition separate them. Every model still hallucinates—even the best invent answers about 9–10% of the time when the documents don't contain one—and per-task cost ranges from $0.003 to $0.36, with price a poor proxy for quality.

Then comes the twist. In 101 cohorts of 100 masked private companies, models rank which will show the strongest momentum over 90 and 180 days, without tools. A heuristic that just sorts by prior headcount growth scores ρ = +0.058. The 13 models that finished all cohorts score between −0.160 and −0.003, every one below the rule (Grok 4.6 completed only 9 cohorts due to timeouts and was excluded). A key reason: models consistently favor larger companies, while smaller ones grew web traffic faster.

The result depends on the target. Scored on headcount growth alone, all models correlate positively and most beat the rule; controlling for size, GPT-5.6 Terra, GPT-6 Sol and others significantly outperform it. That sensitivity is itself the lesson—predicting growth hinges on how you define it.

For investors, the takeaway is practical: use AI for research, screening and synthesis, where gains are immediate. For calling the next breakout, a hiring curve is still more reliable than a model's instinct.