Undermind

Sweet spot: a serious researcher running a small number of high-stakes queries where finding the right 15 papers matters more than the time spent. The 90-second wait stops feeling slow once you realize the alternative is two hours of manual filtering through Google Scholar and Semantic Scholar tabs. Honest concerns. The pricing structure is rough on the casual user. Personal at $19/mo gives you only a handful of deep searches per day, and the free trial is small enough that it is hard to evaluate the product against your real workflow. The product also lives or dies on the quality of its reasoning trace — when the agent picks a paper for the wrong reason, you may waste time reading something genuinely irrelevant before realizing. What to pilot. Run three queries through Undermind that you have already researched manually for a real project. Compare the papers Undermind surfaced against the ones you ended up actually citing. If 70%+ of Undermind's top results overlap with your real bibliography — and especially if it surfaced papers you missed and would have wanted — the subscription pays for itself on a single review. If Undermind misses the canonical papers, your field may be too humanities-heavy or too narrow for the current corpus.

Undermind takes a deliberately contrarian position in AI academic search: it is slow on purpose. Where Consensus and Perplexity return results in seconds, Undermind takes 60–90 seconds per query — and uses that time to run an iterative agent that reads candidate papers, reasons about which ones actually answer your question, follows citations and embeddings to find adjacent work, and returns a small set (typically 10–30) of highly relevant papers with explicit reasoning for why each was chosen. The bet is that for genuinely hard research questions — the kind where keyword search returns 5,000 hits and only 12 are useful — depth beats speed. The agent reasons about query intent, expands the search across synonyms and related concepts, and discards papers that look superficially relevant but do not actually address what you asked. The result is closer to "a research assistant spent 90 seconds on this" than "a search engine returned matches." Audience is narrow and serious: PhD students writing thesis chapters, R&D scientists scoping a new direction, evidence-synthesis professionals starting a review. Casual users will find the wait time annoying and unnecessary. Power users find it transformative for the queries that actually matter. Free trial gives a handful of searches. Personal at $19/mo is the entry tier for individual researchers. Pro at $99/mo lifts search limits substantially for heavy users. Team is custom with shared workspaces. The price is premium because the underlying compute is — every query is genuinely an agent run, not a retrieval call.