Nebius acquired Tavily for $275M in February 2026. This walks the procurement evaluation framework for choosing a search API in mid-2026.
Prerequisites
- Workload definition
- Risk tolerance
- Budget tier
Walkthrough
Step 1: Classify workload
Per-call typed JSON OR long-horizon orchestration?
Text
// Per-call: search → JSON → next step. Orchestration: multi-step research run for 5+ minutes.Step 2: Filter by vendor independence
Pure-play vs acquired-into-bigger-co.
Text
// Pure-play: Scavio, Exa, Serper, Brave Search API, Parallel.
// Acquired/dependent: Tavily (Nebius).
// In litigation: SerpAPI (Google DMCA).Step 3: Match shape to workload
Per-call multi-platform → Scavio. Orchestration → Parallel. Neural similarity → Exa.
Text
// Decision rule per workload.Step 4: Run a 50-query golden set
Compare top-N output quality and latency.
Text
// Score on relevance + latency + cost across candidates.Step 5: Cost-model at expected volume
Project per-month at expected query rate.
Text
// 100K queries/mo: Scavio Project tier; Tavily PAYG; Exa with-content.Step 6: Vendor risk score
Acquisition / lawsuit / dependency.
Text
// Score 1-5 each: vendor-stability, runway, ecosystem-momentum, legal-risk.Step 7: Decide + multi-vendor for production
Default + fallback for resilience.
Text
// 60% default, 30% fallback, 10% legacy. Drop legacy at risk events.Python Example
Python
# Total evaluation: 1-3 days; the win is risk-aware procurement.JavaScript Example
JavaScript
// Evaluation framework is the same.Expected Output
JSON
Documented vendor decision with workload fit, cost projection, vendor-risk rationale, and multi-vendor production setup.