Autonomous Research Tools: Market Map Digest
TL;DR
This cycle the backend deep-research layer consolidated into a priced, benchmarked market: Exa launched Agent Ultra as a metered subagent-swarm API with a $20 default per-run cap, Perplexity retired its Sonar tiers and moved deep research exclusively to the Agent API with per-tool-invocation billing, and publishers (Wiley, Clarivate, EBSCO) published consumer-facing prices for agent access to full text. Meanwhile the verification problem moved from the edges to the center — Exa graded its own benchmarks, and the demand for independent verification is now the clearest opening on the map.
The backend layer now has three priced vendors — and deep research is sold like compute
The destination-vs-backend split hardened into a real market this week: deep research is now a metered API product with published unit economics, not a chat feature. Exa's Agent Ultra, launched September 26, runs a subagent swarm where "Ultra runs typically finish complex tasks in about 30 minutes," metered at standard Agent rates with a default cap of $20 per run, adjustable from $1 to $100 — deep research sold like compute, not like a seat (exa-agent-ultra-deep-research-api+1).
"Sonar billed tokens. The Agent API bills tokens plus per-tool-invocation billing: web_search $2.50, fetch_url $0.50, finance_search $5, people_search $5 per 1,000 invocations." — DEV Community
Perplexity's Sonar retirement (executed September 27 as announced) completes the picture: the preset mapping sonar-deep-research→high, plus a new xhigh tier, means "effort tiers" are now the standard pricing grammar across both Exa and Perplexity — a cost-quality dial rather than a product tier (perplexity-deep-research-consumer-agent). For an orchestrator like Hey, Lefty, this is the procurement moment: three backends (Parallel, Perplexity Agent API, Exa) with three published unit-economics models means routing cheap steps to fast presets and reserving high-step budgets for hard subtasks is now a measurable margin lever (perplexity-deep-research-consumer-agent
).
What to watch: Whether Exa publishes an independent or third-party benchmark response — the thread on its self-graded WANDR claims is open and unresolved.
Publishers picked sides: rent the corpus via MCP, or build the workspace
The data-owning incumbents split into two strategies this cycle, and both now have published price points. Wiley's Scholar Gateway MCP — "first major academic publishers to launch an official MCP server for AI agent access" — covers 8M+ articles with full-text search, and its pricing is the first published consumer-facing price for publisher full-text via MCP: a free tier of 20 results per query and 30 queries per month, with the paid plan at 500 queries for US$15/month (publisher-mcp-licensed-access-layer). Clarivate's Nexus Connect routes ChatGPT and Claude queries into university library collections; EBSCO's AI Exchange connects licensed full text to external platforms including a Perplexity partnership.
"Trust Cards – Show the exact passage used to support each AI-generated insight for easy verification." — HKUST Library LibGuide
Elsevier took the opposite bet: LeapSpace is a full research workspace on its own 16M+ full-text articles and 105M+ Scopus abstracts — the corpus as destination, not rental (publisher-mcp-licensed-access-layer). The structural driver is that full text is the moat while abstracts are the commodity: tools built on Semantic Scholar/OpenAlex inherit the post-2024 abstract restrictions, which is exactly why full-text owners can charge. Note that verification UX — Trust Cards, Wiley's citation-metadata preservation, Clarivate's Nexus Extend citation checking — is converging into table stakes across the layer.
What to watch: Whether Wiley's $15/month anchor pulls other publishers into published consumer pricing or pushes them toward Clarivate-style institutional gateways.
Vertical research tools are winning by embedding into the system of record — and the hosts are embedding back
OpenEvidence's September wave shows the vertical endgame compounding on every front at once: a quiet $250M raise at a $15B valuation, a Memorial Sloan Kettering deal that embeds OE into Epic workflows while OE integrates MSK's OncoKB precision-oncology database, an Anthropic partnership giving physicians in 100 lower-income countries open access, and a pivot into drug development with "the first drug entering clinical trials before year-end" (openevidence-ehr-embed-vertical-distribution).
"MSK is plugging OE into its Epic workflows, and OE is integrating MSK's OncoKB precision oncology database into its engine." — Digital Health Wire
The countermove came from the horizontal side: OpenAI wired Epic — holding data for over 325 million patients — into ChatGPT Health on September 1, read-only, with a Healthcare Public Data plug-in (ClinicalTrials.gov, CMS Coverage, PubMed) and BAA-covered compliant workflows (openevidence-ehr-embed-vertical-distribution). The lesson for the map: institutional embedding, not destination traffic, is where vertical distribution settles — and OpenEvidence's rumored "major system" banning a competitor product over shadow-AI enforcement shows that distribution can be revoked as well as won.
What to watch: Whether OpenEvidence's drug-development pivot repositions it out of the research-tool market entirely — a research company becoming a pharma company changes who it competes with.
Expertise proxies are breaking on both the demand and supply side
The intellectual frame for the whole market arrived this cycle: Sean Goedecke's essay, anchored to the declaration signed by almost five thousand mathematicians including twenty-five Fields medalists, argues that puzzle-solving was a proxy for conceptual progress that AI can now game — "the whole point of those puzzles disappears" (ai-expertise-proxies-goodhart).
"It used to be that you could put a meaty project on your GitHub... But now projects like that are worthless, because everyone just assumes they're vibe-coded." — Sean Goedecke
This is the demand side of the research-tool market: when output volume stops signaling expertise, buyers pay for verification, provenance, and legible method. The twist is that the same Goodhart dynamic has hit the supply side — Exa's WANDR benchmark run "swaps in Exa as the content tool, changes transport logic, and uses gpt-6-luna as the judge," and The Neuron's independent roundup flags the comparisons as vendor claims (exa-agent-ultra-deep-research-api+1). Independent verification is becoming the scarcest commodity in the stack — and the HN debate over whether the QB/WR analogy proves AI is complementary or substitutive shows even the pushback is unresolved (ai-expertise-proxies-goodhart
).
What to watch: Any Aaron Tay-style independent gold-standard test of Exa Agent Ultra's claims — the first one would set the template for procurement-grade evaluation across all three backends.
What surprised us
- The retirement was badly announced by the vendor that sells trust. Perplexity's September 27 Sonar retirement date appears in the blog post and a docs warning banner, but not in the migration guide (which says Sonar "remains supported") nor in the changelog its own FAQ tells developers to watch (perplexity-deep-research-consumer-agent
). For a company selling citation-grounded reliability to developers, that's a telling lapse.
- The citation-array break is a silent killer. The Agent API's response-shape change breaks citation-array parsing — "the thing most source-attribution UIs were built on" — silently, and Perplexity publishes no context-window figure for the new API (perplexity-deep-research-consumer-agent
- Exa's target users confirm the backend thesis precisely. Model providers, financial services, and GTM teams — all workflows where the buyer wants exhaustive recall plus citations, none of which is a chat surface (exa-agent-ultra-deep-research-api
+1). The destination products (Gemini Deep Research, Perplexity's consumer app) and the backend layer are now genuinely different markets.
- A $15B valuation came with "deafening silence." OpenEvidence's round followed $250M at $12B in January and a reported weighing of $200M at $20B in July — "not necessarily a down round, but the silence was deafening" (openevidence-ehr-embed-vertical-distribution
).
Open threads worth a vote
- Independent verification of Exa Agent Ultra's benchmark claims — should the next hunt prioritize independent spot-checks of the WANDR numbers, or wait for Perplexity's official response?