Sixtyfour

Sixtyfour takes one identifier — a name, email, phone number, username, or company name — and deploys AI agents to search the open web, social platforms, public records, and (per the vendor) dark web sources, cross-referencing findings into a structured, cited profile. It ships as a web app (app.sixtyfour.ai), a REST API, and an MCP server that plugs into Claude, Cursor, Codex, and other AI coding tools (docs.sixtyfour.ai/developer-tools/mcp-setup). The company also sells a separate Search API for querying existing people/company records with natural-language or MongoDB-style filters, and a Workflows product for chaining enrichment steps.

Sixtyfour built and open-sourced RECON, a 140-person, 514-field benchmark (github.com/sixtyfour-ai/benchmarks) specifically for the kind of research that requires chaining facts across sources — e.g., find someone's prior startup, then that startup's launch post, then its first customer. In Sixtyfour's own reproducible results (raw JSON at raw.githubusercontent.com/sixtyfour-ai/benchmarks), its 'High' tier scored 71.8% standard accuracy and +56.6% weighted accuracy (correct minus wrong, out of total fields) versus Parallel's best configuration (Ultra 8x) at 44.6%/+31.3%, and out-of-the-box GPT-5.4, Gemini 3.1 Pro, Grok 4.3, and Exa all scored well below both. Sixtyfour High also ran in about half the latency of Parallel Ultra 8x (459s vs 678s p50). This is a vendor-run benchmark, but the scripts, prompts, and scoring methodology are public and re-runnable — that's a meaningfully higher bar of transparency than a marketing claim.

The same benchmark shows Sixtyfour High got 78 of 514 fields wrong (15.2% of attempted answers, precision 82.6%) and missed 67 more. The vendor's own write-up says it plainly: 'Even our best system gets roughly 1 in 10 answers wrong... critical findings should still be verified against primary sources.' For a trust & safety or KYB workflow, a wrong answer that looks confident and cited is the expensive failure mode, not a missing one — Sixtyfour's benchmark is explicit that it weights wrong answers as worse than missing ones for exactly this reason.

Sixtyfour uses a credit system, published in full at docs.sixtyfour.ai/guides/credits-and-pricing (verified 2026-08-27). Plans: Free (64 credits, no rollover), Individual (100 credits/month), Starter (1,200 credits/month, rollover up to 2x monthly cap), and Enterprise (custom). None of the plan pages show a dollar figure for Individual or Starter — the credits-per-plan table has no attached price, so a buyer can see the unit economics (what each operation costs in credits) but not the monthly bill without signing up or contacting sales. Per-operation costs are granular and transparent: a 'Low' tier people-enrichment costs 1 credit, 'Medium' 20 credits, 'High' (full OSINT) 40 credits and is gated to enterprise customers; a professional email lookup costs 0.5 credits, a personal email lookup 2 credits, a phone lookup 3 credits. Credit top-ups cost 3x the plan rate and expire after a year.

Founded 2025, YC Spring 2025 batch (ycombinator.com/companies/sixtyfour), 12-person team per YC's listing (as of 2026-08-27), based in San Francisco. Founders are Saarth Shah (CEO, ex-Whatnot, ex-Deepgram) and Christopher Price (COO, ex-Onyx). Public funding data is thin and inconsistent across trackers: one third-party aggregator (startupintros.com, via Bing search) cites a $500K seed round in April 2025 — we could not independently verify this figure against a primary source (no press release or SEC filing found), so treat it as unconfirmed. The company's own YC launch posts (May 2025 and May 2026) name Cloaked, TRM Labs, Airwallex, Whatnot, Surge, Turing, and Datacurve as customers; these are the vendor's own claims and we did not independently confirm them with the named companies.

Sixtyfour runs a public trust center at compliance.sixtyfour.ai (Vanta-style portal) listing a SOC 2 Type 2 report (request-gated, not open-download), a GDPR-oriented privacy policy, a DPA template, and 16 named subprocessors (AWS, Vercel, Supabase, PostHog, Slack, and others, all US-based per the portal). The published security commitments include AES-256 encryption at rest, TLS in transit, enforced MFA on privileged accounts, daily backups with twice-yearly restore tests, an annual external pen test, and a stated commitment not to train AI models on customer data. Separately, the public privacy policy (version 1.2, published July 23, 2026, at compliance.sixtyfour.ai/en/documents/privacy) states that profile data collected about people (not platform users) is retained for a maximum of 12 months and deleted within 1 month of a deletion request, and lists GDPR Article 15–21 rights plus a CCPA/CPRA section, with a public request form.

We found no G2, Trustpilot, or Capterra listing for Sixtyfour, and no independent press coverage (TechCrunch, mainstream tech press) beyond YC's own Launch YC posts and a handful of funding-tracker aggregator pages (Crunchbase, PitchBook, Tracxn) that mostly restate the company's own YC profile. The benchmark's own limitations section states the RECON dataset (140 people, 514 fields as of the June 2026 run) is not publicly released because it contains real personal data, so outside parties can run the harness against their own comparison set but cannot audit Sixtyfour's exact test set.