The benchmark tested seven providers against GPT-5.6 Luna across quality, cost, and latency. Most comparisons I've seen only look at relevance scores. This one actually weights the three dimensions agents care about: can the model trust the snippets, does the bill stay reasonable at scale, and does the round-trip finish before the user gets impatient.
Parallel took the quality crown. Their results consistently returned the most citation-worthy sources with the least hallucination-inducing noise. For research-heavy agents that need to synthesize across 20+ sources, that gap matters. Exa won on the speed-cost frontier — sub-800ms p95 with pricing that doesn't punish you for deep pagination. Firecrawl surprised me: their extraction quality held up against the bigger names while being the only one with a generous free tier that actually lets you prototype without a credit card.
The also-rans — Tavily, Serper, Bing, and Google Custom Search — each had a fatal flaw. Tavily's latency spiked past 2s on complex queries. Serper's snippet quality degraded noticeably past page 2. Bing and Google CSE both impose rate limits that break parallel agent fan-out patterns unless you negotiate enterprise deals.
What the benchmark doesn't capture: API ergonomics. Parallel's streaming response format lets you start parsing before the full payload lands. Exa's category
filter (news, academic, github, etc.) saves a ton of post-filtering logic. Firecrawl's markdown output is cleaner than anything else I've tried — zero regex cleanup needed before feeding into the context window.
If you're building a coding agent that needs GitHub issues and docs, Exa's category routing is a force multiplier. For general research agents, Parallel's quality edge compounds over multi-hop reasoning. For side projects and MVPs, Firecrawl's free tier removes the "but what if this costs $200/month" hesitation. I've migrated two production pipelines to Parallel + Exa hybrid routing based on query type. The quality delta showed up immediately in eval scores — fewer "I couldn't find reliable sources" failures, more complete answers on the first pass.
Worth pulling the full report if you're evaluating. The raw latency distributions and cost-per-1k-queries tables saved me a week of A/B testing.
Next Stop Losing Home Service Jobs Because Your Follow-Up Takes Four → a practical ChatGPT prompt guide, with plenty of directly applicable cases.